Development and Interpretation of a Deep Learning Method for the Differential Diagnosis of Inflammatory Bowel Disease and Intestinal Tuberculosis

Development and Interpretation of a Deep Learning Method for the Differential Diagnosis of Inflammatory Bowel Disease and Intestinal Tuberculosis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Development and Interpretation of a Deep Learning Method for the Differential Diagnosis of Inflammatory Bowel Disease and Intestinal Tuberculosis Yi Peng, Wei Chen, Shuo Cao, Sha Cheng, Ju Luo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9305072/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 7 You are reading this latest preprint version Abstract Background Distinguishing between inflammatory bowel disease (IBD) and intestinal tuberculosis (ITB) is clinically challenging. The question of whether the deep learning (DL) method can assist in the diagnosis of IBD and ITB remains to be explored. Therefore, this study collected colonoscopy images of IBD and ITB patients, used various DL methods to train and validate differential diagnosis models for the differential diagnosis of IBD and ITB, and introduced various methods to interpret the models. Methods This retrospective study was conducted at the Third Xiangya Hospital and involved data from IBD and ITB patients who were treated from January 2013 to June 2021. A total of 2,612 colonoscopy images from 430 patients were included. Four mainstream DL models were used in this study, including VGG19, InceptionV3, Xception, and ResNet50. Gradient-weighted class activation mapping (Grad-CAM) and traditional medical statistical methods were used for model interpretation. Result Xception performed the best in differentiating between UC and ITB, with an AUC of 0.819, an accuracy of 0.849, a precision of 0.809, a recall of 0.732, and an F1 score of 0.759. Xception performed the best in differentiating between CD and ITB, with an AUC of 0.761, an accuracy of 0.857, a precision of 0.791, a recall of 0.716, and an F1 score of 0.743. Grad-CAM effectively visualized areas of interest, and analysis of the highest-predicted images confirmed that the models recognized valuable endoscopic features for differential diagnosis. Conclusion This study developed two DL models to differentiate between UC and ITB and between CD and ITB, thereby achieving high diagnostic performance and good interpretability. Deep learning Xception Ulcerative colitis Crohn's disease Intestinal tuberculosis Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Inflammatory bowel disease (IBD) is a chronic inflammatory disorder associated with autoimmune disorders, including ulcerative colitis (UC), Crohn's disease (CD), and indeterminate colitis (IC). Intestinal tuberculosis (ITB) is an infection in the gastrointestinal tract caused by Mycobacterium tuberculosis and is characterized by inflammation, granuloma formation, and caseous necrosis [ 1 ]. Clinically, distinguishing between IBD and ITB is challenging, and colonoscopy plays a critical role in diagnosis, management, prognosis, and monitoring. Both domestic and international guidelines recommend colonoscopy as the primary diagnostic and differential tool [ 2 , 3 ]. However, colonoscopy also has several limitations. (1) There is difficulty in differential diagnosis. Although colonoscopy is essential for diagnosing and managing IBD and ITB, distinguishing between these two conditions via colonoscopy alone is challenging. To standardize colonoscopy procedures, the European Crohn's and Colitis Organization (ECCO) developed specific endoscopic descriptors, including loss of vascular patterns, mucosal granularity, erythema, increased friability, bleeding, erosion, ulceration (evaluated via size, shape, depth, and distribution), scars, strictures, polypoid hyperplasia, and cobblestone appearance [ 4 ]. The observation and accurate recording of these features within a short time period during colonoscopy is inherently challenging. (2) There is variability in diagnostic accuracy. Colonoscopy often depends heavily on the physician’s skill and experience. Although colonoscopy is now widely available in community hospitals, diagnoses of IBD and ITB often require expert endoscopists who are typically only available in major tertiary hospitals. Diagnostic capabilities in community hospitals still need improvement. (3) There is variability in positive findings. The quality of colonoscopy relies not only on the physician but also on the patient’s compliance, cooperation, and bowel preparation [ 5 ]. Even the same physician may produce different results for the same patient under varying circumstances. Deep learning (DL), which is a branch of machine learning (ML), relies primarily on convolutional neural networks (CNNs) to extract image features. With the powerful feature extraction capabilities of CNNs, DL has achieved remarkable success in image recognition, including in medical imaging, with models such as VGG, Inception, Xception, and ResNet yielding notable advancements. For example, Cao Z et al. developed a DL model using VGG to distinguish COVID-19 from chest X-rays, with an accuracy of 95.5% [ 6 ]. Schielein MC et al. used Inception to assist in dermatological diagnoses, whereby it achieved high accuracy, sensitivity, and specificity, thus demonstrating that DL can aid in clinical diagnosis and even train dermatology residents [ 7 ]. Shahabi MS et al. utilized Xception to develop an AI system to predict treatment responses for major depression from brain imaging, with an accuracy of 99.3% [ 8 ]. In a previous study, we used ResNet as a base framework to design an AI model that automatically differentiated community-acquired pneumonia from COVID-19 pneumonia with an accuracy of 93.8%, which matched the performance of experienced radiologists [ 9 ]. These studies demonstrate that DL methods are well-suited for complex medical imaging tasks, thus assisting clinicians and enhancing diagnostic skills for less experienced physicians. DL has also shown success in gastroenterology applications. Kim JH et al. developed a DL model based on VGG that uses real-time endoscopy videos to predict the depth of early gastric cancer invasion, whereby it achieved a sensitivity of 82.3%, specificity of 85.8%, and accuracy of 83.7% [ 10 ]. Liu G et al. developed an Inception-based AI model that could differentiate esophageal cancer from precancerous lesions, thus achieving classification accuracies of 94.2%, 82.5%, and 77.1% for normal esophagus, precancerous lesions, and esophageal cancer, respectively [ 11 ]. Hu Y et al. used Xception to create an AI model that identifies lymph node metastasis in gastric cancer via pathology images, and it achieved an accuracy of 97.1%, thus assisting pathologists with preliminary screening [ 12 ]. Our previous study used ResNet to design an AI system for identifying chronic atrophic gastritis and assessing atrophy severity, whereby it achieved accuracies of 89.0% for detection and 77.3% for severity assessment [ 13 ]. These studies confirm that DL models, including VGG, Inception, Xception, and ResNet, are applicable to endoscopic images and gastrointestinal disease diagnosis, thus indicating promising potential for DL in IBD and ITB differentiation via colonoscopy. Although a few studies have explored DL applications in IBD, such as automatic ulcer identification in CD patients [ 14 ], recurrence risk assessment [ 15 ], severity assessment in UC patients [ 16 ], and recurrence risk prediction in UC patients [ 17 ], few studies have focused on differentiating IBD patients from ITB patients. Most studies have prioritized model performance (i.e., diagnostic accuracy), with a limited focus on interpretability and understanding how models make predictions. However, the interpretation of predictions is crucial for translating DL from research to clinical practice. Therefore, this study collected colonoscopy images of IBD and ITB patients, used various DL methods to train and validate differential diagnosis models for the differential diagnosis of IBD and ITB, and introduced various methods to interpret the models. Materials and Methods Subjects and Study Workflow This retrospective study was conducted at the Third Xiangya Hospital and involved data from IBD and ITB patients who were treated from January 2013 to June 2021. All of the patients were diagnosed according to the 2018 IBD diagnostic consensus (Beijing) by two experienced gastroenterologists who considered clinical presentations, laboratory findings, endoscopy, pathology, and treatment responses. The study included 168 UC, 195 CD, and 67 ITB patients, each of whom underwent colonoscopy. The main workflow involved (1) establishing a dataset by collecting colonoscopy images for the three diseases; (2) training and validating DL-assisted diagnostic models; and (3) interpreting the model results. See Fig. 1 for the study workflow. Establishing the Colonoscopy Image Dataset Eligible cases had abnormal images extracted from their colonoscopy results. The dataset comprised 2,612 colonoscopy images from 430 cases, which included UC (967 images), CD (1,184 images), and ITB (461 images) images. The images were divided into training, validation, and test sets at an 8:1:1 ratio. See Table 1 for the image distribution. Table 1 Dataset of colonoscopy images Colonoscopy image dataset Number of colonoscopy cases Number of colonoscopy images Training set (80%) Validation set (10%) Testing set (10%) ITB 67 461 369 46 46 CD 195 1184 948 118 118 UC 168 967 773 97 97 Deep Learning Methods Four mainstream DL models were used in this study, including VGG19, InceptionV3, Xception, and ResNet50. VGG19, which was developed by Karen Simonyan et al., uses small (3×3) convolutional layers and performs excellently in both localization and classification tasks [ 18 ]. Inception, which was developed by Christian Szegedy et al., has a carefully crafted structure to increase network depth and width while maintaining computational efficiency [ 19 ]. Xception improves upon Inception by decoupling channels and spatial correlations [ 20 ]. ResNet, which was developed by Kaiming He et al., incorporates residual functions and skip connections, thus effectively mitigating gradient vanishing issues and increasing model depth and performance [ 21 ]. Interpretability Methods As DL models increasingly impact various aspects of daily life, model interpretability is crucial for gaining users' trust, especially in health care. To improve the understanding of model predictions, two interpretability methods were used: (1) the gradient-weighted class activation mapping (Grad-CAM) method; and (2) the traditional medical statistical method, which analyzes the microscopic features in predictive images. Grad-CAM can help us in analyzing the network's attention area for a certain category. We can subsequently analyze whether the network has learned the correct features or information through the network's attention area. Grad-CAM is essentially the reverse use of the deep learning method. It uses the calculation formula to backpropagate the predicted score of the predicted category and then uses the layer information that was backpropagated to the feature to calculate the importance of each channel in the feature layer. The data of each channel in the feature layer are subsequently weighted and summed. Finally, Grad-CAM is obtained by activating the function, and a heatmap is drawn. In this manner, we can visualize the feature information that is recognized by the DL model [ 22 ]. In traditional medical statistical methods, the specific procedural steps are as follows. First, the model will give a predicted value of 0 to 1 when making predictions. The closer the predicted value is to 1, the greater the possibility that the model predicts the image as a certain disease. We collected 15 images of each disease from the test set in descending order of the predicted value. These images are considered by the model to best represent the disease. Second, we separately analyzed the microscopic features of each image. The collected microscopic features refer to the consensus opinion of European IBD endoscopy [ 4 ]. Finally, we counted the endoscopic features of each disease on the endoscopic images and analyzed whether these features were different. Thus, the endoscopic features that the model uses to distinguish diseases can be inferred. Experiment Parameters and Evaluation Metrics All of the DL experiments were conducted on a cloud server ( https://www.matpool.com/ ). The utilized GPU was an NVIDIA Tesla K80 GPU×4 with an environment of Python 3.5, CUDA 10.0, cuDNN 7.6, TensorFlow 1.13.1, Keras 2.2, and Ubuntu 18.04. The parameters included batch size (4), epochs (100), and learning rate (0.0003). Tenfold cross-validation was implemented, with early stopping being utilized to prevent overfitting. Dataset division followed an 8:1:1 split for training, validation, and testing. The codes, models, and parameters are publicly available at https://github.com/philiplaw1984/IBD/ . The evaluation values that were used to evaluate the performance of the model in this study included the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F-1 score. Statistical Analysis and Ethics Statistical analysis was performed via SPSS 22.0 (IBM SPSS Statistics, Armonk, NY, USA). Categorical variables were analyzed via the chi-square test or Fisher’s exact test (if sample sizes were small), with two-tailed P values < 0.05 considered to be statistically significant. This study was approved by the Ethics Committee of Clinical Research at the Third Xiangya Hospital of Central South University (Approval No. 22272) and was conducted in accordance with the Declaration of Helsinki. For the retrospective collection of stored colonoscopy images from hospital patients, the Ethics Committee granted a waiver for informed consent. Results Performance of the DL-assisted Diagnostic Models Differentiating between UC and ITB: The AUC results for the VGG19, ResNet50, InceptionV3, and Xception models were 0.776, 0.744, 0.638, and 0.819, respectively, as shown in Fig. 2 . Additional evaluation metrics such as accuracy, precision, recall, and F1 score for each model are provided in Table 2 . Among the DL models that were tested, Xception performed the best, with an AUC of 0.819, an accuracy of 0.849, a precision of 0.809, a recall of 0.732, and an F1 score of 0.759. To minimize sampling error during training and testing, tenfold cross-validation was conducted on the Xception model, yielding average values of 0.838 ± 0.042 for accuracy, 0.803 ± 0.071 for precision, 0.708 ± 0.073 for recall, and 0.732 ± 0.073 for the F1 score, as shown in Table 3 . Table 2 Results of four different deep learning methods that help in distinguishing UC from ITB VGG19 AUC Accuracy Precision Recall F1 Score 0.776 0.833 0.812 0.674 0.705 ResNet50 0.744 0.785 0.692 0.607 0.621 InceptionV3 0.638 0.785 0.729 0.558 0.552 Xception 0.819 0.849 0.809 0.732 0.759 Table 3 Tenfold cross-validation results for Xception, which help to distinguish UC from ITB 1 Accuracy Precision Recall F1 Score 0.849 0.809 0.732 0.759 2 0.817 0.772 0.651 0.677 3 0.849 0.801 0.744 0.766 4 0.833 0.833 0.662 0.694 5 0.812 0.748 0.736 0.741 6 0.801 0.732 0.629 0.649 7 0.841 0.871 0.667 0.702 8 0.841 0.809 0.703 0.734 9 0.793 0.704 0.672 0.685 10 0.942 0.947 0.883 0.910 0.838 ± 0.042 0.803 ± 0.071 0.708 ± 0.073 0.732 ± 0.073 Differentiating between CD and ITB: The AUC values for the VGG19, ResNet50, InceptionV3, and Xception models were 0.701, 0.746, 0.704, and 0.761, respectively, as depicted in Fig. 3 . Additional model performance metrics, including accuracy, precision, recall, and F1 score, are presented in Table 4 . The Xception model again achieved the best results, with an AUC of 0.761, an accuracy of 0.857, a precision of 0.791, a recall of 0.716, and an F1 score of 0.743. Tenfold cross-validation for Xception yielded average accuracy, precision, recall, and F1 scores of 0.890 ± 0.058, 0.835 ± 0.098, 0.791 ± 0.118, and 0.806 ± 0.110, respectively, as detailed in Table 5 . Table 4 Results of four different deep learning methods that help to distinguish CD from ITB VGG19 AUC Accuracy Precision Recall F1 Score 0.701 0.816 0.704 0.625 0.645 ResNet50 0.746 0.836 0.754 0.664 0.690 InceptionV3 0.704 0.829 0.746 0.634 0.659 Xception 0.761 0.857 0.791 0.716 0.743 Table 5 Tenfold cross-validation results for Xception, which help in distinguishing CD from ITB 1 Accuracy Precision Recall F1 Score 0.857 0.791 0.716 0.743 2 0.871 0.864 0.698 0.741 3 0.911 0.863 0.853 0.858 4 0.945 0.923 0.901 0.911 5 0.843 0.762 0.694 0.718 6 0.850 0.764 0.751 0.757 7 0.972 0.947 0.970 0.958 8 0.979 0.987 0.948 0.966 9 0.864 0.791 0.759 0.773 10 0.812 0.667 0.625 0.640 0.890 ± 0.058 0.835 ± 0.098 0.791 ± 0.118 0.806 ± 0.110 Interpretability of the DL-assisted Diagnostic Model The Grad-CAM method was used in this study to visualize the regions of interest that were identified by the DL models. Figure 4 shows the following information: (1) input image, which is the original colonoscopy image being examined; (2) heatmap image, which represents the areas of interest generated by the Grad-CAM method, thus directly identifying the relative locations of the lesions; and (3) Grad-CAM image, which is an overlay of the input image and heatmap. This visualization allows for an intuitive understanding of how the DL model makes decisions. To further understand which endoscopic features the DL model recognized for each condition (UC, CD, and ITB), the 15 images with the highest predicted values in the test set were analyzed for distinguishing endoscopic features. For UC and ITB differentiation, (1) among the 15 UC images, 9 images showed increased mucosal fragility and bleeding compared with 2 images in the ITB images (60.0% vs. 13.3%, respectively; P = 0.021); and (2) six of the 15 ITB images showed ring ulcers, with none observed in the UC images (40.0% vs. 0.0%, respectively; P = 0.017), as detailed in Table 6 . For CD and ITB differentiation, (1) eight of the 15 CD images exhibited longitudinal ulcers, with no lesions observed in the ITB images (53.3% vs. 0.0%, respectively; P = 0.002); (2) eleven of the 15 CD images exhibited a cobblestone appearance, with no lesions observed in the ITB images (73.3% vs. 0.0%, respectively; P < 0.001); and (3) five of the 15 ITB images showed scarring changes, with no scarring observed in the CD images (33.3% vs. 0.0%, respectively; P = 0.042), as detailed in Table 7 . Table 6 Image feature analysis (UC vs. ITB) Increased mucosal fragility and bleeding UC(N = 15) ITB(N = 15) P 9(60.0%) 2(13.3%) 0.021 Ring ulcer 0 6(40.0%) 0.017 Table 7 Image feature analysis (CD vs. ITB) Longitudinal ulcer CD(N = 15) ITB(N = 15) P 8(53.3%) 0 0.002 Cobblestone 11(73.3%) 0 <0.001 Scar 0 5(33.3%) 0.042 Discussion This study developed three DL models for the differential diagnosis of UC, ITB, and CD based on colonoscopy images and utilized Grad-CAM and the statistical analysis of endoscopic image features to explain the models. The DL-assisted diagnostic models exhibited strong diagnostic performance with interpretability, which aids physicians in diagnosis and provides a rationale for clinical decisions. Xception outperformed other DL models in distinguishing between ITB and CD, as well as between ITB and UC, which is consistent with research showing its effectiveness in detecting gastric ulcers, rectal cancer, and pulmonary nodules [ 23 – 25 ]. Xception, which is a CNN architecture based on depthwise separable convolution, includes 36 convolution layers for feature extraction, and it focuses on characteristics such as color, texture, shape, and spatial relationships. The loss of mucosal vascular texture, which is a common feature in colonoscopy, is typical of texture features, whereas mucosal redness and bleeding are color features. Ulcers, scars, strictures, polypoid growth, and cobblestone changes include multiple image features, thus allowing the strong image feature extraction capability of Xception to distinguish ITB from CD and ITB from UC. This study achieved an AUC of 0.761 for the model differentiating ITB and CD. A Korean study using colonoscopy images of 2,123 CD patients and 1,642 ITB patients reported an AUC of 0.785, which was slightly greater than that in our study; this result was likely due to a larger dataset [ 26 ]. This finding highlights the potential for improved performance with increased sample size, although such gains may be marginal because models can reach a point in which added images provide diminishing returns. In clinical applications, DL diagnostic models must demonstrate high performance and interpretability, as models must substantiate their predictions with scientific and medical bases for clinicians to trust them with patient health. DL models that provide a diagnostic rationale can be effectively integrated into clinical workflows. Grad-CAM has been applied to various types of medical images, such as MRI for multiple sclerosis and CT for COVID-19 diagnosis [ 27 , 28 ]. This study demonstrated that using Grad-CAM with colonoscopy images can effectively localize lesions, thus indicating its suitability for different imaging modalities. Although the DL model can learn the features of the image well, the Grad-CAM method can explain the region of interest of the DL model (i.e., the localization of the lesion via the DL model). However, the image features that the DL model has learned still need to be further explained with medical professional knowledge. When considering the differentiation between UC and ITB, after analyzing the images with the highest predictive value of the model, this study revealed that the most important features of the UC group were mucosal hemorrhage and increased fragility, whereas the most important feature of the ITB group was circular ulcers. Among them, mucosal hemorrhage and increased fragility are common manifestations of UC patients under colonoscopy [ 2 , 29 ]. Circular ulcers are among the diagnostic manifestations of ITB patients under colonoscopy [ 2 ]. For the differentiation between CD patients and ITB patients, a previous study from South Korea used clinical data from 40 CD patients and 40 ITB patients to investigate disease prediction models. Predictive factors for CD include lesions, longitudinal ulcers, aphthous ulcers, and paving stone changes. The predictive factors for tuberculosis involve the possession of fewer than 4 segments of the affected colon, a dilated ileocecal valve, annular ulcers, scars, and polypoid hyperplasia [ 30 ]. After the images with the highest predictive value of the model were analyzed, the most important features of the CD group were longitudinal ulcers and paving stone changes, whereas the most important features of the ITB group were stenosis, scars, and polypoid hyperplasia. These findings are essentially consistent with literature reports and clinical experience. Therefore, by analyzing the image with the highest predicted value, it can be confirmed that the features learned by the DL differential diagnosis model that was developed in this study are all microscopic features with good diagnostic efficiency. The use of the DL differential diagnosis model should be safe and reliable. This study also has several limitations. First, we focused only on single-image differentiation without incorporating patient-level case differentiation. Additionally, the limited number of ITB colonoscopy images somewhat constrained the model’s ability to identify ITB. Future research will involve the collection of more images of intestinal tuberculosis to further enhance the performance of the DL model. Conclusion Via colonoscopy image data, this study developed two DL models to differentiate UC and ITB and CD and ITB, thus achieving high diagnostic performance. Grad-CAM effectively visualized areas of interest, thus assisting physicians in understanding the model’s diagnostic approach. The analysis of the highest-predicted images confirmed that the models recognized endoscopic features that are valuable for differential diagnosis. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of Clinical Research at the Third Xiangya Hospital of Central South University (Approval No. 22272) and was conducted in accordance with the Declaration of Helsinki. For the retrospective collection of stored colonoscopy images from hospital patients, the Ethics Committee granted a waiver for informed consent. Consent for publication Not applicable. Author contribution statement Yi Peng: Methodology, Formal analysis and Writing - Original Draft Wei Chen: Writing - Original Draft and Funding acquisition. Shuo Cao: Data Curation and Validation Sha Cheng: Data Curation and Visualization Yi Peng: Software and Funding acquisition. Ju Luo: Conceptualization, Supervision, and Writing - Review & Editing and Funding acquisition. Funding statement The Science and Technology Program of Hunan Province ‌(2024JJ9054, 2025JJ80565); Scientific Research Project of the Hunan Education Department for Young Scholars (24B0414). Data availability statement The codes, models, and parameters are publicly available at https://github.com/philiplaw1984/IBD/. Declaration of interest statement The authors declare that they have no competing interests. Additional information No additional information is available for this paper. Acknowledgments None References Waljee AK, Sauder K, Patel A, et al. Machine Learning Algorithms for Objective Remission and Clinical Outcomes with Thiopurines. J Crohns Colitis. 2017;11(7):801–10. 10.1093/ecco-jcc/jjx014 . Inflammatory Bowel Disease Group, Chinese Society of Gastroenterology, Chinese Medical Association. Chinese consensus on diagnosis and treatment in inflammatory bowel disease(2018, Beijing). Chin J Digestion. 2018;38(5):292–311. 10.3760/cma.j.issn.0254-1432.2018.05.002 . Lamb CA, Kennedy NA, Raine T et al. British Society of Gastroenterology consensus guidelines on the management of inflammatory bowel disease in adults [published correction appears in Gut. 2021;70(4):1. doi: 10.1136/gutjnl-2019-318484corr1]. Gut . 2019;68(Suppl 3):s1-s106. 10.1136/gutjnl-2019-318484 Annese V, Daperno M, Rutter MD, et al. European evidence based consensus for endoscopy in inflammatory bowel disease. J Crohns Colitis. 2013;7(12):982–1018. 10.1016/j.crohns.2013.09.016 . Shuangzhe Y, Hailong C. Quality control in colonoscopy. Chin J Digestion. 2016;33(3):137–9. 10.3760/cma.j.issn.1007-5232.2016.03.001 . Cao Z, Huang J, He X, et al. BND-VGG-19: A deep learning algorithm for COVID-19 identification utilizing X-ray images. Knowl Based Syst. 2022;258:110040. 10.1016/j.knosys.2022.110040 . Schielein MC, Christl J, Sitaru S, et al. Outlier detection in dermatology: Performance of different convolutional neural networks for binary classification of inflammatory skin diseases. J Eur Acad Dermatol Venereol. 2023;37(5):1071–9. 10.1111/jdv.18853 . Shahabi MS, Shalbaf A, Nobakhsh B, et al. Attention-Based Convolutional Recurrent Deep Neural Networks for the Prediction of Response to Repetitive Transcranial Magnetic Stimulation for Major Depressive Disorder. Int J Neural Syst. 2023;33(2):2350007. 10.1142/S0129065723500077 . Luo J, Sun Y, Chi J, et al. A novel deep learning-based method for COVID-19 pneumonia detection from CT images. BMC Med Inf Decis Mak. 2022;22(1):284. 10.1186/s12911-022-02022-1 . Published 2022 Nov 2. Kim JH, Oh SI, Han SY, et al. An Optimal Artificial Intelligence System for Real-Time Endoscopic Prediction of Invasion Depth in Early Gastric Cancer. Cancers (Basel). 2022;14(23):6000. 10.3390/cancers14236000 . Published 2022 Dec 5. Liu G, Hua J, Wu Z, et al. Automatic classification of esophageal lesions in endoscopic images using a convolutional neural network. Ann Transl Med. 2020;8(7):486. 10.21037/atm.2020.03.24 . Hu Y, Su F, Dong K, et al. Deep learning system for lymph node quantification and metastatic cancer identification from whole-slide pathology images. Gastric Cancer. 2021;24(4):868–77. 10.1007/s10120-021-01158-9 . Luo J, Cao S, Ding N, et al. A deep learning method to assist with chronic atrophic gastritis diagnosis using white light images. Dig Liver Dis. 2022;54(11):1513–9. 10.1016/j.dld.2022.04.025 . Klang E, Barash Y, Margalit RY, et al. Deep learning algorithms for automated detection of Crohn's disease ulcers by video capsule endoscopy. Gastrointest Endosc. 2020;91(3):606–e6132. 10.1016/j.gie.2019.11.012 . Kiyokawa H, Abe M, Matsui T, et al. Deep Learning Analysis of Histologic Images from Intestinal Specimen Reveals Adipocyte Shrinkage and Mast Cell Infiltration to Predict Postoperative Crohn Disease. Am J Pathol. 2022;192(6):904–16. 10.1016/j.ajpath.2022.03.006 . Fan Y, Mu R, Xu H, et al. Novel deep learning-based computer-aided diagnosis system for predicting inflammatory activity in ulcerative colitis. Gastrointest Endosc. 2023;97(2):335–46. 10.1016/j.gie.2022.08.015 . Ohara J, Nemoto T, Maeda Y, Ogata N, Kudo SE, Yamochi T. Deep learning-based automated quantification of goblet cell mucus using histological images as a predictor of clinical relapse of ulcerative colitis with endoscopic remission. J Gastroenterol. 2022;57(12):962–70. 10.1007/s00535-022-01924-1 . Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image. Recognition[J] Comput Sci. 2014. 10.48550/arXiv.1409.1556 . Szegedy C, Liu W, Jia Y, et al. Going Deeper Convolutions[J] IEEE Comput Soc. 2014. 10.1109/CVPR.2015.7298594 . Chollet F. Xception: Deep Learning with Depthwise Separable Convolutions[J].IEEE, 2017. 10.1109/cvpr.2017.195 He K, Zhang X, Ren S et al. Deep Residual Learning for Image Recognition[J]. https://doi.org/10.48550/arXiv.1512.03385 Selvaraju RR, Cogswell M, Das A et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization[C]//IEEE International Conference on Computer Vision. IEEE, 2017. 10.1109/ICCV.2017.74 Liu Y, Zhang L, Hao Z et al. An xception model based on residual attention mechanism for the classification of benign and malignant gastric ulcers. Sci Rep . 2022;12(1):15365. Published 2022 Sep 13. 10.1038/s41598-022-19639-x Yi S, Wei Y, Luo X et al. Diagnosis of rectal cancer based on the Xception-MS network. Phys Med Biol . 2022;67(19): 10.1088/1361–6560/ac8f11 . Published 2022 Sep 19. doi:10.1088/1361-6560/ac8f11. Li D, Yuan S, Yao G. Classification of lung nodules based on the DCA-Xception network. J Xray Sci Technol. 2022;30(5):993–1008. 10.3233/XST-221219 . Kim JM, Kang JG, Kim S, Cheon JH. Deep-learning system for real-time differentiation between Crohn's disease, intestinal Behçet's disease, and intestinal tuberculosis. J Gastroenterol Hepatol. 2021;36(8):2141–8. 10.1111/jgh.15433 . Zhang Y, Hong D, McClement D, et al. Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging. J Neurosci Methods. 2021;353:109098. 10.1016/j.jneumeth.2021.109098 . Panwar H, Gupta PK, Siddiqui MK, et al. A deep learning and grad-CAM based color visualization approach for fast detection of COVID-19 cases using chest X-ray and CT-Scan images. Chaos Solitons Fractals. 2020;140:110190. 10.1016/j.chaos.2020.110190 . Pera A, Bellando P, Caldera D, et al. Colonoscopy in inflammatory bowel disease. Diagnostic accuracy and proposal of an endoscopic score. Gastroenterology. 1987;92(1):181–5. Bae JH, Park SH, Ye BD, et al. Development and Validation of a Novel Prediction Model for Differential Diagnosis Between Crohn's Disease and Intestinal Tuberculosis. Inflamm Bowel Dis. 2017;23(9):1614–23. 10.1097/MIB.0000000000001162 . Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 15 May, 2026 Reviewers agreed at journal 07 May, 2026 Reviewers invited by journal 07 May, 2026 Editor invited by journal 13 Apr, 2026 Editor assigned by journal 04 Apr, 2026 Submission checks completed at journal 04 Apr, 2026 First submitted to journal 02 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9305072","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":641501658,"identity":"b471b029-36ff-4176-aa36-4a4def913e6a","order_by":0,"name":"Yi Peng","email":"","orcid":"","institution":"The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China","correspondingAuthor":false,"prefix":"","firstName":"Yi","middleName":"","lastName":"Peng","suffix":""},{"id":641501659,"identity":"1a7548d4-b335-4529-a904-5756ac1adac8","order_by":1,"name":"Wei Chen","email":"","orcid":"","institution":"The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Chen","suffix":""},{"id":641501661,"identity":"21e25393-fa32-47d4-983c-c0e6f710bd10","order_by":2,"name":"Shuo Cao","email":"","orcid":"","institution":"Third Xiangya Hospital, Central South University","correspondingAuthor":false,"prefix":"","firstName":"Shuo","middleName":"","lastName":"Cao","suffix":""},{"id":641501662,"identity":"20ee2cba-8d31-49d1-8490-a5f158b4c295","order_by":3,"name":"Sha Cheng","email":"","orcid":"","institution":"Third Xiangya Hospital, Central South University","correspondingAuthor":false,"prefix":"","firstName":"Sha","middleName":"","lastName":"Cheng","suffix":""},{"id":641501663,"identity":"7db8d90e-122b-4ac4-a29c-c1948adbe5e3","order_by":4,"name":"Ju Luo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1UlEQVRIiWNgGAWjYBACPiA2YKiQA1I8YAHGBkJa2MBazhiTqIWBsY0kLRLJB4p55xkkNrCfPbqZh8FGdsMB5mcP8GtJSzDm3QbUwpOXdpuHIc14wwE2cwP8WnIMgFr+JDYw5JgBtRxO3HCAh02CsJY5QFv434C0/CdWSwNQiwTYlgNEaOF5lmA455iBcZvEG7ObcwySjWceZjPDq4WfPfmYwZsaA9l+/hyzG28q7GT7jjc/w6sFZBEofBzbwGwQk5mAepCSB0DCnrC6UTAKRsEoGLEAAKmKQIMPpPtuAAAAAElFTkSuQmCC","orcid":"","institution":"The Fifth Affiliated Hospital of Anhui Medical University","correspondingAuthor":true,"prefix":"","firstName":"Ju","middleName":"","lastName":"Luo","suffix":""}],"badges":[],"createdAt":"2026-04-02 15:38:43","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9305072/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9305072/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109446020,"identity":"a278a2c4-03b3-4361-9ca8-5ddfc5b88ec9","added_by":"auto","created_at":"2026-05-18 08:17:32","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":320056,"visible":true,"origin":"","legend":"\u003cp\u003eStudy workflow\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9305072/v1/252f0be5683991084e7e9495.png"},{"id":109760344,"identity":"7d5b022c-7612-47bb-aed4-eedf8d38335d","added_by":"auto","created_at":"2026-05-22 07:28:34","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":44744,"visible":true,"origin":"","legend":"\u003cp\u003eAUCs of VGG19, ResNet50, InceptionV3, and Xception for Differentiation UC and ITB\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9305072/v1/af13922d5cfbd6b742091da1.png"},{"id":109760283,"identity":"e2159565-2c7b-4565-bb6d-6093cf7ee06f","added_by":"auto","created_at":"2026-05-22 07:28:25","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":43437,"visible":true,"origin":"","legend":"\u003cp\u003eAUCs of VGG19, ResNet50, InceptionV3, and Xception for Differentiation CD and ITB\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9305072/v1/7b1e83845acaed5aa0c61a58.png"},{"id":109799494,"identity":"8759f484-71e2-4e28-b2c7-f6aeea1ffe67","added_by":"auto","created_at":"2026-05-22 15:29:51","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1729061,"visible":true,"origin":"","legend":"\u003cp\u003eThe Grad-CAM method for visualizing the regions of interest that were identified by the DL models\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9305072/v1/3ba72197068526cb32bd3507.png"},{"id":109907083,"identity":"7c45061e-5eb1-4f9b-8361-2e93f536fdaa","added_by":"auto","created_at":"2026-05-25 06:41:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2242398,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9305072/v1/b217a748-d33a-49a0-ab5c-1dca6c26f08e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and Interpretation of a Deep Learning Method for the Differential Diagnosis of Inflammatory Bowel Disease and Intestinal Tuberculosis","fulltext":[{"header":"Introduction","content":"\u003cp\u003eInflammatory bowel disease (IBD) is a chronic inflammatory disorder associated with autoimmune disorders, including ulcerative colitis (UC), Crohn's disease (CD), and indeterminate colitis (IC). Intestinal tuberculosis (ITB) is an infection in the gastrointestinal tract caused by \u003cem\u003eMycobacterium tuberculosis\u003c/em\u003e and is characterized by inflammation, granuloma formation, and caseous necrosis [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Clinically, distinguishing between IBD and ITB is challenging, and colonoscopy plays a critical role in diagnosis, management, prognosis, and monitoring. Both domestic and international guidelines recommend colonoscopy as the primary diagnostic and differential tool [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eHowever, colonoscopy also has several limitations. (1) There is difficulty in differential diagnosis. Although colonoscopy is essential for diagnosing and managing IBD and ITB, distinguishing between these two conditions via colonoscopy alone is challenging. To standardize colonoscopy procedures, the European Crohn's and Colitis Organization (ECCO) developed specific endoscopic descriptors, including loss of vascular patterns, mucosal granularity, erythema, increased friability, bleeding, erosion, ulceration (evaluated via size, shape, depth, and distribution), scars, strictures, polypoid hyperplasia, and cobblestone appearance [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The observation and accurate recording of these features within a short time period during colonoscopy is inherently challenging. (2) There is variability in diagnostic accuracy. Colonoscopy often depends heavily on the physician\u0026rsquo;s skill and experience. Although colonoscopy is now widely available in community hospitals, diagnoses of IBD and ITB often require expert endoscopists who are typically only available in major tertiary hospitals. Diagnostic capabilities in community hospitals still need improvement. (3) There is variability in positive findings. The quality of colonoscopy relies not only on the physician but also on the patient\u0026rsquo;s compliance, cooperation, and bowel preparation [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Even the same physician may produce different results for the same patient under varying circumstances.\u003c/p\u003e \u003cp\u003eDeep learning (DL), which is a branch of machine learning (ML), relies primarily on convolutional neural networks (CNNs) to extract image features. With the powerful feature extraction capabilities of CNNs, DL has achieved remarkable success in image recognition, including in medical imaging, with models such as VGG, Inception, Xception, and ResNet yielding notable advancements. For example, Cao Z et al. developed a DL model using VGG to distinguish COVID-19 from chest X-rays, with an accuracy of 95.5% [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Schielein MC et al. used Inception to assist in dermatological diagnoses, whereby it achieved high accuracy, sensitivity, and specificity, thus demonstrating that DL can aid in clinical diagnosis and even train dermatology residents [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Shahabi MS et al. utilized Xception to develop an AI system to predict treatment responses for major depression from brain imaging, with an accuracy of 99.3% [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. In a previous study, we used ResNet as a base framework to design an AI model that automatically differentiated community-acquired pneumonia from COVID-19 pneumonia with an accuracy of 93.8%, which matched the performance of experienced radiologists [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. These studies demonstrate that DL methods are well-suited for complex medical imaging tasks, thus assisting clinicians and enhancing diagnostic skills for less experienced physicians.\u003c/p\u003e \u003cp\u003eDL has also shown success in gastroenterology applications. Kim JH et al. developed a DL model based on VGG that uses real-time endoscopy videos to predict the depth of early gastric cancer invasion, whereby it achieved a sensitivity of 82.3%, specificity of 85.8%, and accuracy of 83.7% [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Liu G et al. developed an Inception-based AI model that could differentiate esophageal cancer from precancerous lesions, thus achieving classification accuracies of 94.2%, 82.5%, and 77.1% for normal esophagus, precancerous lesions, and esophageal cancer, respectively [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Hu Y et al. used Xception to create an AI model that identifies lymph node metastasis in gastric cancer via pathology images, and it achieved an accuracy of 97.1%, thus assisting pathologists with preliminary screening [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Our previous study used ResNet to design an AI system for identifying chronic atrophic gastritis and assessing atrophy severity, whereby it achieved accuracies of 89.0% for detection and 77.3% for severity assessment [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. These studies confirm that DL models, including VGG, Inception, Xception, and ResNet, are applicable to endoscopic images and gastrointestinal disease diagnosis, thus indicating promising potential for DL in IBD and ITB differentiation via colonoscopy.\u003c/p\u003e \u003cp\u003eAlthough a few studies have explored DL applications in IBD, such as automatic ulcer identification in CD patients [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], recurrence risk assessment [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], severity assessment in UC patients [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], and recurrence risk prediction in UC patients [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], few studies have focused on differentiating IBD patients from ITB patients. Most studies have prioritized model performance (i.e., diagnostic accuracy), with a limited focus on interpretability and understanding how models make predictions. However, the interpretation of predictions is crucial for translating DL from research to clinical practice.\u003c/p\u003e \u003cp\u003eTherefore, this study collected colonoscopy images of IBD and ITB patients, used various DL methods to train and validate differential diagnosis models for the differential diagnosis of IBD and ITB, and introduced various methods to interpret the models.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eSubjects and Study Workflow\u003c/h2\u003e \u003cp\u003eThis retrospective study was conducted at the Third Xiangya Hospital and involved data from IBD and ITB patients who were treated from January 2013 to June 2021. All of the patients were diagnosed according to the 2018 IBD diagnostic consensus (Beijing) by two experienced gastroenterologists who considered clinical presentations, laboratory findings, endoscopy, pathology, and treatment responses. The study included 168 UC, 195 CD, and 67 ITB patients, each of whom underwent colonoscopy.\u003c/p\u003e \u003cp\u003eThe main workflow involved (1) establishing a dataset by collecting colonoscopy images for the three diseases; (2) training and validating DL-assisted diagnostic models; and (3) interpreting the model results. See Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e for the study workflow.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eEstablishing the Colonoscopy Image Dataset\u003c/h3\u003e\n\u003cp\u003eEligible cases had abnormal images extracted from their colonoscopy results. The dataset comprised 2,612 colonoscopy images from 430 cases, which included UC (967 images), CD (1,184 images), and ITB (461 images) images. The images were divided into training, validation, and test sets at an 8:1:1 ratio. See Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e for the image distribution.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDataset of colonoscopy images\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003eColonoscopy image dataset\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of colonoscopy cases\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of colonoscopy images\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTraining set (80%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eValidation set (10%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTesting set\u003c/p\u003e \u003cp\u003e(10%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eITB\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e461\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e369\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCD\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e195\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1184\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e948\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e118\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e118\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eUC\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e168\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e967\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e773\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eDeep Learning Methods\u003c/h3\u003e\n\u003cp\u003eFour mainstream DL models were used in this study, including VGG19, InceptionV3, Xception, and ResNet50. VGG19, which was developed by Karen Simonyan et al., uses small (3\u0026times;3) convolutional layers and performs excellently in both localization and classification tasks [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Inception, which was developed by Christian Szegedy et al., has a carefully crafted structure to increase network depth and width while maintaining computational efficiency [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Xception improves upon Inception by decoupling channels and spatial correlations [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. ResNet, which was developed by Kaiming He et al., incorporates residual functions and skip connections, thus effectively mitigating gradient vanishing issues and increasing model depth and performance [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eInterpretability Methods\u003c/h3\u003e\n\u003cp\u003eAs DL models increasingly impact various aspects of daily life, model interpretability is crucial for gaining users' trust, especially in health care. To improve the understanding of model predictions, two interpretability methods were used: (1) the gradient-weighted class activation mapping (Grad-CAM) method; and (2) the traditional medical statistical method, which analyzes the microscopic features in predictive images.\u003c/p\u003e \u003cp\u003eGrad-CAM can help us in analyzing the network's attention area for a certain category. We can subsequently analyze whether the network has learned the correct features or information through the network's attention area. Grad-CAM is essentially the reverse use of the deep learning method. It uses the calculation formula to backpropagate the predicted score of the predicted category and then uses the layer information that was backpropagated to the feature to calculate the importance of each channel in the feature layer. The data of each channel in the feature layer are subsequently weighted and summed. Finally, Grad-CAM is obtained by activating the function, and a heatmap is drawn. In this manner, we can visualize the feature information that is recognized by the DL model [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn traditional medical statistical methods, the specific procedural steps are as follows. First, the model will give a predicted value of 0 to 1 when making predictions. The closer the predicted value is to 1, the greater the possibility that the model predicts the image as a certain disease. We collected 15 images of each disease from the test set in descending order of the predicted value. These images are considered by the model to best represent the disease. Second, we separately analyzed the microscopic features of each image. The collected microscopic features refer to the consensus opinion of European IBD endoscopy [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Finally, we counted the endoscopic features of each disease on the endoscopic images and analyzed whether these features were different. Thus, the endoscopic features that the model uses to distinguish diseases can be inferred.\u003c/p\u003e\n\u003ch3\u003eExperiment Parameters and Evaluation Metrics\u003c/h3\u003e\n\u003cp\u003eAll of the DL experiments were conducted on a cloud server (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.matpool.com/\u003c/span\u003e\u003cspan address=\"https://www.matpool.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The utilized GPU was an NVIDIA Tesla K80 GPU\u0026times;4 with an environment of Python 3.5, CUDA 10.0, cuDNN 7.6, TensorFlow 1.13.1, Keras 2.2, and Ubuntu 18.04. The parameters included batch size (4), epochs (100), and learning rate (0.0003). Tenfold cross-validation was implemented, with early stopping being utilized to prevent overfitting. Dataset division followed an 8:1:1 split for training, validation, and testing. The codes, models, and parameters are publicly available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/philiplaw1984/IBD/\u003c/span\u003e\u003cspan address=\"https://github.com/philiplaw1984/IBD/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe evaluation values that were used to evaluate the performance of the model in this study included the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F-1 score.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis and Ethics\u003c/h2\u003e \u003cp\u003eStatistical analysis was performed via SPSS 22.0 (IBM SPSS Statistics, Armonk, NY, USA). Categorical variables were analyzed via the chi-square test or Fisher\u0026rsquo;s exact test (if sample sizes were small), with two-tailed P values\u0026thinsp;\u0026lt;\u0026thinsp;0.05 considered to be statistically significant.\u003c/p\u003e \u003cp\u003eThis study was approved by the Ethics Committee of Clinical Research at the Third Xiangya Hospital of Central South University (Approval No. 22272) and was conducted in accordance with the Declaration of Helsinki. For the retrospective collection of stored colonoscopy images from hospital patients, the Ethics Committee granted a waiver for informed consent.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003ePerformance of the DL-assisted Diagnostic Models\u003c/h2\u003e \u003cp\u003eDifferentiating between UC and ITB: The AUC results for the VGG19, ResNet50, InceptionV3, and Xception models were 0.776, 0.744, 0.638, and 0.819, respectively, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Additional evaluation metrics such as accuracy, precision, recall, and F1 score for each model are provided in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Among the DL models that were tested, Xception performed the best, with an AUC of 0.819, an accuracy of 0.849, a precision of 0.809, a recall of 0.732, and an F1 score of 0.759. To minimize sampling error during training and testing, tenfold cross-validation was conducted on the Xception model, yielding average values of 0.838\u0026thinsp;\u0026plusmn;\u0026thinsp;0.042 for accuracy, 0.803\u0026thinsp;\u0026plusmn;\u0026thinsp;0.071 for precision, 0.708\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073 for recall, and 0.732\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073 for the F1 score, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of four different deep learning methods that help in distinguishing UC from ITB\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eVGG19\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.776\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.812\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.674\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.705\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.744\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.785\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.692\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.607\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.621\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInceptionV3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.638\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.785\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.729\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.558\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.552\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eXception\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.819\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.849\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.809\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.732\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e0.759\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTenfold cross-validation results for Xception, which help to distinguish UC from ITB\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.849\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.809\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.732\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.759\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.817\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.772\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.651\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.677\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.801\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.744\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.766\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.662\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.694\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.812\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.748\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.736\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.741\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.801\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.732\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.629\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.649\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.841\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.871\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.702\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.841\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.809\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.703\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.734\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.793\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.704\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.672\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.685\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.942\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.947\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.883\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.910\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.838\u0026thinsp;\u0026plusmn;\u0026thinsp;0.042\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.803\u0026thinsp;\u0026plusmn;\u0026thinsp;0.071\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.708\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.732\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eDifferentiating between CD and ITB: The AUC values for the VGG19, ResNet50, InceptionV3, and Xception models were 0.701, 0.746, 0.704, and 0.761, respectively, as depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Additional model performance metrics, including accuracy, precision, recall, and F1 score, are presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The Xception model again achieved the best results, with an AUC of 0.761, an accuracy of 0.857, a precision of 0.791, a recall of 0.716, and an F1 score of 0.743. Tenfold cross-validation for Xception yielded average accuracy, precision, recall, and F1 scores of 0.890\u0026thinsp;\u0026plusmn;\u0026thinsp;0.058, 0.835\u0026thinsp;\u0026plusmn;\u0026thinsp;0.098, 0.791\u0026thinsp;\u0026plusmn;\u0026thinsp;0.118, and 0.806\u0026thinsp;\u0026plusmn;\u0026thinsp;0.110, respectively, as detailed in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of four different deep learning methods that help to distinguish CD from ITB\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eVGG19\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.701\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.816\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.704\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.625\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.645\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.746\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.836\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.754\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.664\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.690\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInceptionV3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.704\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.829\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.746\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.634\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.659\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eXception\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.761\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.857\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.791\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.716\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e0.743\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTenfold cross-validation results for Xception, which help in distinguishing CD from ITB\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.857\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.791\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.716\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.743\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.871\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.698\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.741\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.911\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.863\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.853\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.858\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.945\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.923\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.901\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.911\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.843\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.762\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.694\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.718\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.850\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.764\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.751\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.757\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.972\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.947\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.970\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.958\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.979\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.987\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.948\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.966\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.791\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.759\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.773\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.812\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.640\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.890\u0026thinsp;\u0026plusmn;\u0026thinsp;0.058\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.835\u0026thinsp;\u0026plusmn;\u0026thinsp;0.098\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.791\u0026thinsp;\u0026plusmn;\u0026thinsp;0.118\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.806\u0026thinsp;\u0026plusmn;\u0026thinsp;0.110\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eInterpretability of the DL-assisted Diagnostic Model\u003c/h2\u003e \u003cp\u003eThe Grad-CAM method was used in this study to visualize the regions of interest that were identified by the DL models. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows the following information: (1) input image, which is the original colonoscopy image being examined; (2) heatmap image, which represents the areas of interest generated by the Grad-CAM method, thus directly identifying the relative locations of the lesions; and (3) Grad-CAM image, which is an overlay of the input image and heatmap. This visualization allows for an intuitive understanding of how the DL model makes decisions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo further understand which endoscopic features the DL model recognized for each condition (UC, CD, and ITB), the 15 images with the highest predicted values in the test set were analyzed for distinguishing endoscopic features. For UC and ITB differentiation, (1) among the 15 UC images, 9 images showed increased mucosal fragility and bleeding compared with 2 images in the ITB images (60.0% vs. 13.3%, respectively; P\u0026thinsp;=\u0026thinsp;0.021); and (2) six of the 15 ITB images showed ring ulcers, with none observed in the UC images (40.0% vs. 0.0%, respectively; P\u0026thinsp;=\u0026thinsp;0.017), as detailed in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. For CD and ITB differentiation, (1) eight of the 15 CD images exhibited longitudinal ulcers, with no lesions observed in the ITB images (53.3% vs. 0.0%, respectively; P\u0026thinsp;=\u0026thinsp;0.002); (2) eleven of the 15 CD images exhibited a cobblestone appearance, with no lesions observed in the ITB images (73.3% vs. 0.0%, respectively; P\u0026thinsp;\u0026lt;\u0026thinsp;0.001); and (3) five of the 15 ITB images showed scarring changes, with no scarring observed in the CD images (33.3% vs. 0.0%, respectively; P\u0026thinsp;=\u0026thinsp;0.042), as detailed in Table\u0026nbsp;\u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eImage feature analysis (UC vs. ITB)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eIncreased mucosal fragility and bleeding\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUC(N\u0026thinsp;=\u0026thinsp;15)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eITB(N\u0026thinsp;=\u0026thinsp;15)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9(60.0%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2(13.3%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.021\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRing ulcer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6(40.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.017\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eImage feature analysis (CD vs. ITB)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eLongitudinal ulcer\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCD(N\u0026thinsp;=\u0026thinsp;15)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eITB(N\u0026thinsp;=\u0026thinsp;15)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8(53.3%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCobblestone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11(73.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScar\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5(33.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.042\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study developed three DL models for the differential diagnosis of UC, ITB, and CD based on colonoscopy images and utilized Grad-CAM and the statistical analysis of endoscopic image features to explain the models. The DL-assisted diagnostic models exhibited strong diagnostic performance with interpretability, which aids physicians in diagnosis and provides a rationale for clinical decisions.\u003c/p\u003e \u003cp\u003eXception outperformed other DL models in distinguishing between ITB and CD, as well as between ITB and UC, which is consistent with research showing its effectiveness in detecting gastric ulcers, rectal cancer, and pulmonary nodules [\u003cspan additionalcitationids=\"CR24\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Xception, which is a CNN architecture based on depthwise separable convolution, includes 36 convolution layers for feature extraction, and it focuses on characteristics such as color, texture, shape, and spatial relationships. The loss of mucosal vascular texture, which is a common feature in colonoscopy, is typical of texture features, whereas mucosal redness and bleeding are color features. Ulcers, scars, strictures, polypoid growth, and cobblestone changes include multiple image features, thus allowing the strong image feature extraction capability of Xception to distinguish ITB from CD and ITB from UC.\u003c/p\u003e \u003cp\u003eThis study achieved an AUC of 0.761 for the model differentiating ITB and CD. A Korean study using colonoscopy images of 2,123 CD patients and 1,642 ITB patients reported an AUC of 0.785, which was slightly greater than that in our study; this result was likely due to a larger dataset [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. This finding highlights the potential for improved performance with increased sample size, although such gains may be marginal because models can reach a point in which added images provide diminishing returns.\u003c/p\u003e \u003cp\u003eIn clinical applications, DL diagnostic models must demonstrate high performance and interpretability, as models must substantiate their predictions with scientific and medical bases for clinicians to trust them with patient health. DL models that provide a diagnostic rationale can be effectively integrated into clinical workflows.\u003c/p\u003e \u003cp\u003eGrad-CAM has been applied to various types of medical images, such as MRI for multiple sclerosis and CT for COVID-19 diagnosis [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. This study demonstrated that using Grad-CAM with colonoscopy images can effectively localize lesions, thus indicating its suitability for different imaging modalities.\u003c/p\u003e \u003cp\u003eAlthough the DL model can learn the features of the image well, the Grad-CAM method can explain the region of interest of the DL model (i.e., the localization of the lesion via the DL model). However, the image features that the DL model has learned still need to be further explained with medical professional knowledge.\u003c/p\u003e \u003cp\u003eWhen considering the differentiation between UC and ITB, after analyzing the images with the highest predictive value of the model, this study revealed that the most important features of the UC group were mucosal hemorrhage and increased fragility, whereas the most important feature of the ITB group was circular ulcers. Among them, mucosal hemorrhage and increased fragility are common manifestations of UC patients under colonoscopy [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Circular ulcers are among the diagnostic manifestations of ITB patients under colonoscopy [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor the differentiation between CD patients and ITB patients, a previous study from South Korea used clinical data from 40 CD patients and 40 ITB patients to investigate disease prediction models. Predictive factors for CD include lesions, longitudinal ulcers, aphthous ulcers, and paving stone changes. The predictive factors for tuberculosis involve the possession of fewer than 4 segments of the affected colon, a dilated ileocecal valve, annular ulcers, scars, and polypoid hyperplasia [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAfter the images with the highest predictive value of the model were analyzed, the most important features of the CD group were longitudinal ulcers and paving stone changes, whereas the most important features of the ITB group were stenosis, scars, and polypoid hyperplasia. These findings are essentially consistent with literature reports and clinical experience. Therefore, by analyzing the image with the highest predicted value, it can be confirmed that the features learned by the DL differential diagnosis model that was developed in this study are all microscopic features with good diagnostic efficiency. The use of the DL differential diagnosis model should be safe and reliable.\u003c/p\u003e \u003cp\u003eThis study also has several limitations. First, we focused only on single-image differentiation without incorporating patient-level case differentiation. Additionally, the limited number of ITB colonoscopy images somewhat constrained the model\u0026rsquo;s ability to identify ITB. Future research will involve the collection of more images of intestinal tuberculosis to further enhance the performance of the DL model.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eVia colonoscopy image data, this study developed two DL models to differentiate UC and ITB and CD and ITB, thus achieving high diagnostic performance. Grad-CAM effectively visualized areas of interest, thus assisting physicians in understanding the model\u0026rsquo;s diagnostic approach. The analysis of the highest-predicted images confirmed that the models recognized endoscopic features that are valuable for differential diagnosis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee of Clinical Research at the Third Xiangya Hospital of Central South University (Approval No. 22272) and was conducted in accordance with the Declaration of Helsinki. For the retrospective collection of stored colonoscopy images from hospital patients, the Ethics Committee granted a waiver for informed consent.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contribution statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eYi Peng: Methodology,\u0026nbsp;Formal analysis and Writing - Original Draft\u003c/p\u003e\n\u003cp\u003eWei Chen: Writing - Original Draft and Funding acquisition.\u003c/p\u003e\n\u003cp\u003eShuo Cao: Data Curation and Validation\u003c/p\u003e\n\u003cp\u003eSha Cheng: Data Curation and Visualization\u003c/p\u003e\n\u003cp\u003eYi Peng: Software and Funding acquisition.\u003c/p\u003e\n\u003cp\u003eJu Luo: Conceptualization, Supervision, and Writing - Review \u0026amp; Editing and Funding acquisition.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Science and Technology Program of Hunan Province \u0026zwnj;(2024JJ9054, 2025JJ80565); Scientific Research Project of the Hunan Education Department for Young Scholars (24B0414).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe codes, models, and parameters are publicly available at https://github.com/philiplaw1984/IBD/.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of interest statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo additional information is available for this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNone\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWaljee AK, Sauder K, Patel A, et al. Machine Learning Algorithms for Objective Remission and Clinical Outcomes with Thiopurines. J Crohns Colitis. 2017;11(7):801\u0026ndash;10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/ecco-jcc/jjx014\u003c/span\u003e\u003cspan address=\"10.1093/ecco-jcc/jjx014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInflammatory Bowel Disease Group, Chinese Society of Gastroenterology, Chinese Medical Association. Chinese consensus on diagnosis and treatment in inflammatory bowel disease(2018, Beijing). Chin J Digestion. 2018;38(5):292\u0026ndash;311. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3760/cma.j.issn.0254-1432.2018.05.002\u003c/span\u003e\u003cspan address=\"10.3760/cma.j.issn.0254-1432.2018.05.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLamb CA, Kennedy NA, Raine T et al. British Society of Gastroenterology consensus guidelines on the management of inflammatory bowel disease in adults [published correction appears in Gut. 2021;70(4):1. doi: 10.1136/gutjnl-2019-318484corr1]. \u003cem\u003eGut\u003c/em\u003e. 2019;68(Suppl 3):s1-s106. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/gutjnl-2019-318484\u003c/span\u003e\u003cspan address=\"10.1136/gutjnl-2019-318484\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnnese V, Daperno M, Rutter MD, et al. European evidence based consensus for endoscopy in inflammatory bowel disease. J Crohns Colitis. 2013;7(12):982\u0026ndash;1018. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.crohns.2013.09.016\u003c/span\u003e\u003cspan address=\"10.1016/j.crohns.2013.09.016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShuangzhe Y, Hailong C. Quality control in colonoscopy. Chin J Digestion. 2016;33(3):137\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3760/cma.j.issn.1007-5232.2016.03.001\u003c/span\u003e\u003cspan address=\"10.3760/cma.j.issn.1007-5232.2016.03.001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao Z, Huang J, He X, et al. BND-VGG-19: A deep learning algorithm for COVID-19 identification utilizing X-ray images. Knowl Based Syst. 2022;258:110040. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.knosys.2022.110040\u003c/span\u003e\u003cspan address=\"10.1016/j.knosys.2022.110040\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchielein MC, Christl J, Sitaru S, et al. Outlier detection in dermatology: Performance of different convolutional neural networks for binary classification of inflammatory skin diseases. J Eur Acad Dermatol Venereol. 2023;37(5):1071\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/jdv.18853\u003c/span\u003e\u003cspan address=\"10.1111/jdv.18853\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShahabi MS, Shalbaf A, Nobakhsh B, et al. Attention-Based Convolutional Recurrent Deep Neural Networks for the Prediction of Response to Repetitive Transcranial Magnetic Stimulation for Major Depressive Disorder. Int J Neural Syst. 2023;33(2):2350007. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1142/S0129065723500077\u003c/span\u003e\u003cspan address=\"10.1142/S0129065723500077\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuo J, Sun Y, Chi J, et al. A novel deep learning-based method for COVID-19 pneumonia detection from CT images. BMC Med Inf Decis Mak. 2022;22(1):284. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12911-022-02022-1\u003c/span\u003e\u003cspan address=\"10.1186/s12911-022-02022-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2022 Nov 2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim JH, Oh SI, Han SY, et al. An Optimal Artificial Intelligence System for Real-Time Endoscopic Prediction of Invasion Depth in Early Gastric Cancer. Cancers (Basel). 2022;14(23):6000. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/cancers14236000\u003c/span\u003e\u003cspan address=\"10.3390/cancers14236000\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2022 Dec 5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu G, Hua J, Wu Z, et al. Automatic classification of esophageal lesions in endoscopic images using a convolutional neural network. Ann Transl Med. 2020;8(7):486. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21037/atm.2020.03.24\u003c/span\u003e\u003cspan address=\"10.21037/atm.2020.03.24\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu Y, Su F, Dong K, et al. Deep learning system for lymph node quantification and metastatic cancer identification from whole-slide pathology images. Gastric Cancer. 2021;24(4):868\u0026ndash;77. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s10120-021-01158-9\u003c/span\u003e\u003cspan address=\"10.1007/s10120-021-01158-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuo J, Cao S, Ding N, et al. A deep learning method to assist with chronic atrophic gastritis diagnosis using white light images. Dig Liver Dis. 2022;54(11):1513\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.dld.2022.04.025\u003c/span\u003e\u003cspan address=\"10.1016/j.dld.2022.04.025\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKlang E, Barash Y, Margalit RY, et al. Deep learning algorithms for automated detection of Crohn's disease ulcers by video capsule endoscopy. Gastrointest Endosc. 2020;91(3):606\u0026ndash;e6132. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.gie.2019.11.012\u003c/span\u003e\u003cspan address=\"10.1016/j.gie.2019.11.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKiyokawa H, Abe M, Matsui T, et al. Deep Learning Analysis of Histologic Images from Intestinal Specimen Reveals Adipocyte Shrinkage and Mast Cell Infiltration to Predict Postoperative Crohn Disease. Am J Pathol. 2022;192(6):904\u0026ndash;16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ajpath.2022.03.006\u003c/span\u003e\u003cspan address=\"10.1016/j.ajpath.2022.03.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFan Y, Mu R, Xu H, et al. Novel deep learning-based computer-aided diagnosis system for predicting inflammatory activity in ulcerative colitis. Gastrointest Endosc. 2023;97(2):335\u0026ndash;46. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.gie.2022.08.015\u003c/span\u003e\u003cspan address=\"10.1016/j.gie.2022.08.015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOhara J, Nemoto T, Maeda Y, Ogata N, Kudo SE, Yamochi T. Deep learning-based automated quantification of goblet cell mucus using histological images as a predictor of clinical relapse of ulcerative colitis with endoscopic remission. J Gastroenterol. 2022;57(12):962\u0026ndash;70. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00535-022-01924-1\u003c/span\u003e\u003cspan address=\"10.1007/s00535-022-01924-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image. Recognition[J] Comput Sci. 2014. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.1409.1556\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1409.1556\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSzegedy C, Liu W, Jia Y, et al. Going Deeper Convolutions[J] IEEE Comput Soc. 2014. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/CVPR.2015.7298594\u003c/span\u003e\u003cspan address=\"10.1109/CVPR.2015.7298594\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChollet F. Xception: Deep Learning with Depthwise Separable Convolutions[J].IEEE, 2017.\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/cvpr.2017.195\u003c/span\u003e\u003cspan address=\"10.1109/cvpr.2017.195\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S et al. Deep Residual Learning for Image Recognition[J]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.1512.03385\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1512.03385\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSelvaraju RR, Cogswell M, Das A et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization[C]//IEEE International Conference on Computer Vision. IEEE, 2017.\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ICCV.2017.74\u003c/span\u003e\u003cspan address=\"10.1109/ICCV.2017.74\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Zhang L, Hao Z et al. An xception model based on residual attention mechanism for the classification of benign and malignant gastric ulcers. \u003cem\u003eSci Rep\u003c/em\u003e. 2022;12(1):15365. Published 2022 Sep 13. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-022-19639-x\u003c/span\u003e\u003cspan address=\"10.1038/s41598-022-19639-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYi S, Wei Y, Luo X et al. Diagnosis of rectal cancer based on the Xception-MS network. \u003cem\u003ePhys Med Biol\u003c/em\u003e. 2022;67(19):\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1088/1361\u0026ndash;6560/ac8f11\u003c/span\u003e\u003cspan address=\"10.1088/1361\u0026ndash;6560/ac8f11\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2022 Sep 19. doi:10.1088/1361-6560/ac8f11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi D, Yuan S, Yao G. Classification of lung nodules based on the DCA-Xception network. J Xray Sci Technol. 2022;30(5):993\u0026ndash;1008. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3233/XST-221219\u003c/span\u003e\u003cspan address=\"10.3233/XST-221219\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim JM, Kang JG, Kim S, Cheon JH. Deep-learning system for real-time differentiation between Crohn's disease, intestinal Beh\u0026ccedil;et's disease, and intestinal tuberculosis. J Gastroenterol Hepatol. 2021;36(8):2141\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/jgh.15433\u003c/span\u003e\u003cspan address=\"10.1111/jgh.15433\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Hong D, McClement D, et al. Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging. J Neurosci Methods. 2021;353:109098. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jneumeth.2021.109098\u003c/span\u003e\u003cspan address=\"10.1016/j.jneumeth.2021.109098\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePanwar H, Gupta PK, Siddiqui MK, et al. A deep learning and grad-CAM based color visualization approach for fast detection of COVID-19 cases using chest X-ray and CT-Scan images. Chaos Solitons Fractals. 2020;140:110190. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.chaos.2020.110190\u003c/span\u003e\u003cspan address=\"10.1016/j.chaos.2020.110190\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePera A, Bellando P, Caldera D, et al. Colonoscopy in inflammatory bowel disease. Diagnostic accuracy and proposal of an endoscopic score. Gastroenterology. 1987;92(1):181\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBae JH, Park SH, Ye BD, et al. Development and Validation of a Novel Prediction Model for Differential Diagnosis Between Crohn's Disease and Intestinal Tuberculosis. Inflamm Bowel Dis. 2017;23(9):1614\u0026ndash;23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/MIB.0000000000001162\u003c/span\u003e\u003cspan address=\"10.1097/MIB.0000000000001162\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Deep learning, Xception, Ulcerative colitis, Crohn's disease, Intestinal tuberculosis","lastPublishedDoi":"10.21203/rs.3.rs-9305072/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9305072/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eDistinguishing between inflammatory bowel disease (IBD) and intestinal tuberculosis (ITB) is clinically challenging. The question of whether the deep learning (DL) method can assist in the diagnosis of IBD and ITB remains to be explored. Therefore, this study collected colonoscopy images of IBD and ITB patients, used various DL methods to train and validate differential diagnosis models for the differential diagnosis of IBD and ITB, and introduced various methods to interpret the models.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis retrospective study was conducted at the Third Xiangya Hospital and involved data from IBD and ITB patients who were treated from January 2013 to June 2021. A total of 2,612 colonoscopy images from 430 patients were included. Four mainstream DL models were used in this study, including VGG19, InceptionV3, Xception, and ResNet50. Gradient-weighted class activation mapping (Grad-CAM) and traditional medical statistical methods were used for model interpretation.\u003c/p\u003e\u003ch2\u003eResult\u003c/h2\u003e \u003cp\u003eXception performed the best in differentiating between UC and ITB, with an AUC of 0.819, an accuracy of 0.849, a precision of 0.809, a recall of 0.732, and an F1 score of 0.759. Xception performed the best in differentiating between CD and ITB, with an AUC of 0.761, an accuracy of 0.857, a precision of 0.791, a recall of 0.716, and an F1 score of 0.743. Grad-CAM effectively visualized areas of interest, and analysis of the highest-predicted images confirmed that the models recognized valuable endoscopic features for differential diagnosis.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThis study developed two DL models to differentiate between UC and ITB and between CD and ITB, thereby achieving high diagnostic performance and good interpretability.\u003c/p\u003e","manuscriptTitle":"Development and Interpretation of a Deep Learning Method for the Differential Diagnosis of Inflammatory Bowel Disease and Intestinal Tuberculosis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-18 08:17:23","doi":"10.21203/rs.3.rs-9305072/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-15T06:35:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"236671522434818542702720365854956903836","date":"2026-05-08T01:19:35+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-05-07T10:58:48+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-13T05:48:43+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-04T11:16:30+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-04T11:15:48+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2026-04-02T15:25:32+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e08b3d1d-4150-4e4c-a470-e88591dffd63","owner":[],"postedDate":"May 18th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-15T06:35:29+00:00","index":54,"fulltext":""},{"type":"reviewerAgreed","content":"236671522434818542702720365854956903836","date":"2026-05-08T01:19:35+00:00","index":35,"fulltext":""},{"type":"reviewersInvited","content":"30","date":"2026-05-07T10:58:48+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-18T08:17:23+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-18 08:17:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9305072","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9305072","identity":"rs-9305072","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}