Using the DCJM deep-learning model to diagnosis drug-resistant lymph node tuberculosis based on ultrasound images: A Multicenter Study

Using the DCJM deep-learning model to diagnosis drug-resistant lymph node tuberculosis based on ultrasound images: A Multicenter Study | 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 Article Using the DCJM deep-learning model to diagnosis drug-resistant lymph node tuberculosis based on ultrasound images: A Multicenter Study Peijun Chen, Jialei Luo, Xinyi Yan, Ying Zhang, Wenzhi Zhang, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5358428/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Our objective was to develop a deep learning model based on grey-scale ultrasound (GUS) images for predicting whether lymph node tuberculosis (LNTB) of the neck is drug resistant. The GUS images of 297 cases of cervical LNTB confirmed to be drug-resistant or sensitive by laboratory examination in three hospitals were retrospectively collected. A target detection-image classification joint learning method (DCJM) model combining target detection and image classification was constructed from the training set, and the diagnostic efficacy of the DCJM model was evaluated by the data from the internal validation set, Test A and Test B. We used mean average precision (mAP) to assess the accuracy of target detection in the DCJM model, The mAP_0.5 and mAP_0.5:0.95 of the DCJM model for LNTB detection were 0.995 and 0.897, respectively. The area under the curve (AUC) of this model in the training set, validation set, Test A, and Test B were 0.992 (95% CI, 0.972-1.000), 0.851 (95% CI, 0.733–0.948), 0.727 (95% CI, 0.488–0.924), and 0.777 (95% CI, 0.644-0.900), respectively. The DCJM model has a strong detection function as well as a good predictive value for drug-resistant LNTB, providing valuable information for individualized treatment decisions. Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Ultrasound Deep learning Lymph node tuberculosis Drug resistance Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Drug-resistant tuberculosis (TB) remains a threat to public health, with 410,000 (3.9%) multidrug/rifampicin-resistant TB patients globally in 2022, accounting for approximately 3.3% of new TB patients and 17% of retreatment patients 1 . The estimated number of multidrug/rifampicin-resistant TB patients in China in 2022 is 30,000, and the rate of extrapulmonary TB resistance is higher than that of single pulmonary TB, making diagnostic and treatment evaluations difficult, and patients take longer (usually more than two years) and cost more to treat 2 . One of the main challenges in controlling multidrug-resistant TB lies in the difficulty of diagnosing drug resistance in patients suspected of having TB, especially in first-time patients. Therefore, the diagnosis and treatment of drug-resistant TB is an urgent challenge in TB prevention and control. Lymph node tuberculosis (LNTB), as a common form of extrapulmonary TB, accounts for 35% of all extrapulmonary TB cases 3 , 4 . LNTB is still treated with systemic anti-tuberculosis drug chemotherapy, which usually involves oral administration of multiple anti-tuberculosis drugs for up to 6–9 months 5 . With the proportion of drug-resistant TB increasing year by year, inadequate detection of drug-resistant TB in pathogenetically positive patients still exists, making subsequent treatment difficult. Early assessment of drug resistance in patients is a prerequisite for timely adjustment of treatment regimens. The main method of obtaining specimens for assessing LNTB drug resistance detection is ultrasound-guided lymph node biopsy, but it is invasive as well as at risk of various potential complications 5 . Therefore, it is urgent to explore a non-invasive screening method that can help improve the detection rate of drug-resistant LNTB. In the past few years, deep learning methods have made significant progress in the field of medical image analysis. Deep learning can significantly improve the recognition and prediction of medical images. It has achieved good performance in diagnosing thyroid, breast and liver diseases 6 – 8 . However, the application of deep learning in lymph node imaging is still relatively rare compared to other fields. Currently, a few studies have reported that deep learning based on ultrasound images of breast and thyroid nodes has been used to predict whether the associated draining lymph nodes are metastatic or not 9 , 10 . Lee et al. 11 collected 812 ultrasound images of cervical lymph nodes that had been pathologically confirmed to develop a computerized system with a view to help improve the accuracy of the diagnosis of lymph node metastasis, with an accuracy of 83%. In lymph node research, artificial intelligence (AI) has been compared to diagnostically experienced physicians and found to have similar diagnostic accuracy with adequate training 12 . There are fewer reports of imaging studies on drug-resistant LNTB, and in this study, we attempted to construct a prediction model for drug-resistant LNTB in the neck by using a deep learning method based on grey-scale ultrasound (GUS) images to explore the value of the model in predicting drug-resistant LNTB. Materials and methods Ethics statement The study was approved by the Medical Ethics Committee of Hangzhou Red Cross Hospital, and written informed consent was obtained from each participant (ApprovaI NO.: [2023] Review No. (046)). Patient enrollment Independent cohort data from three hospitals in China were used in this study. LNTB GUS images from Hangzhou Red Cross Hospital between January 2018 and December 2022 were first retrospectively collected for inclusion in the study, and were randomly assigned proportionally to the training set (70%) and validation set (30%). Lymph node ultrasound images from Infectious Disease Hospital of Heilongjiang Province (Test A) and Kunming Third People's Hospital (Test B) between August 2022 and December 2022 were also collected as the external test set. Baseline characteristics of participants, including age, gender, lymph node status, and histological type. All data were anonymized to protect participant privacy. Patient inclusion criteria were as follows (1) tissue or pus culture and drug susceptibility testing techniques (liquid drug susceptibility testing), or molecular biology rapid resistance testing techniques (including linear probe, Gene-Xpert MTB/RIF assay, and other techniques) were used as diagnostic criteria for LNTB resistance 13 . (2) The general information of the patients was complete. (3) Clear GUS images with distinguishable lesions manually outlined. Exclusion criteria included (1) Unsatisfactory quality of ultrasound images. (2) Lymph nodes that showed rupture phenomenon or were too large to be shown completely. (3) Inability to identify the patient's anti-tuberculosis drug resistance or sensitivity results. (4) Previous history of anti-tuberculosis treatment. GUS image and clinical data acquisition All patients were placed in supine or lateral position for ultrasound examination. Region-by-region scanning was performed to look for suspected enlarged lymph nodes, and the largest lymph nodes were selected for observation. Two radiologists with 5 years of working experience in lymph node ultrasonography were responsible for acquiring the image data, and the lymph node images were saved in DICOM format. The images of each patient were obtained from the PACS database in DICOM format, and were organized based on the case number to obtain the corresponding examination results of the patient. A sheet of GUS static images of the largest longitudinal section of the target lymph node of each patient was selected for subsequent deep learning analysis. The corresponding examination results of the patients were obtained from the clinical case system, including tissue or pus culture, liquid drug susceptibility testing, Gene-Xpert MTB/RIF and whole genome sequencing were used as diagnostic criteria for LNTB resistance. Development and validation of deep learning model As the lymph node GUS image data has different disease classifications, data corruption, abnormal data distribution, and missing key information of the data, the data needs to be pre-processed when analyzing the data. It included format conversion, data screening and cleaning, data effective part extraction, data normalization, and lymph node GUS image enhancement. In addition by using small sample expansion techniques, the training set was expanded by rotating, flipping, scaling and panning the images, and generative adversarial networks. The combined use of these techniques helps to solve the problem of limited samples of lymph node ultrasound images and improve the performance of deep learning models. We proposed a target detection-image classification joint learning method (DCJM) model that combined target detection and image classification (Fig. 1 ). Firstly, the target detection model (YOLOv5 model) was used to detect the key regions of the input ultrasound image, and the high confidence and intersection of union (IoU) were set as the judgement conditions to improve the accuracy of the key region identification, to exclude most of the interfering factors and the influence of the wrong region, and then to classify the extracted key regions. At the same time, in order to avoid the problem that the key regions were not identified due to the high threshold conditions, and thus cannot participate in the classification without predicting the results, the DCJM model used the second image classification model to make auxiliary judgments to optimize the final overall classification results of the joint two-way. Adopting two-way detection, dynamically adjusting the allocation of prediction weights according to the final diagnostic efficacy results of the dataset, and combining the two-way prediction results for judgement, provides a new solution for solving the problem of misdiagnosis and omission of disease diagnosis. To better explain the prediction logic of the YOLOv5 model, we used the gradient-weighted class activation mapping (Grad-CAM) method to visualize the most indicative image regions of the cervical lymph nodes by filtering the feature heatmap from the last convolutional layer 14 . Additionally, we used the t-distributed stochastic neighbor embedding (t-SNE) method to illustrate the overall prediction effect by converting the representation of the last layer of the deep neural network before the prediction node of each image in the test dataset into a shifted or non-shifted colour coding. Experimental environment configuration: Hardware configuration: 12th Gen Intel(R), Core(TM), i9-12900KF (CPU), NVIDIA GeForce RTX 3090 (GPU), 64G (Memory). Software configuration: Ubuntu 22.04, Python 3.9, Pytorch 1.12.0, CUDA 11.3. Statistical Analysis Precision, recall, F1 score, and, mean average precision (mAP) were used in the DCJM model to assess the performance of lymph node detection. IoU is used to measure the degree of overlap between the "predicted border" and the "actual border". mAP 0.5 is the average AP of all the images under each category when the IoU is 0.5. mAP 0.5:0.95 is the average AP of all the images under each category when the mAP threshold is from 50–95%, and then average these ten values. 50% to a mAP threshold of 95%, with 5% intervals, 10 mAP values were obtained, and then these ten values were averaged. Area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were used to assess the predictive ability of the DCJM model for drug-resistant LNTB. The statistical analysis was performed using SPSS software version 25.0. The continuous variables in this study were presented as mean ± SD. All statistical significance levels were set as two-sided with p < 0.05. Results Patient characteristics The model training and internal validation set dataset was obtained from the Hangzhou Red Cross Hospital, and among the 92 patients with cervical LNTB in the drug-resistant group, there were 34 males and 58 females, with an age range of 19–79 years (33.8 ± 15.1 years); among the 142 patients with cervical LNTB in the sensitive group, there were 61 males and 81 females, with an age range of 17–87 years (41.2 ± 18.3 years). There were no significant differences between the two groups of patients in terms of age and gender (all p > 0.05). External test cohort A (Test A) consisted of 30 patients, which contained 7 resistant and 23 sensitive cases. There were 14 males and 16 females with an age range of 22–72 years (35.9 ± 13.1 years). External test cohort B (Test B) consisted of 33 patients, which contained 9 resistant and 24 sensitive cases. There were 12 males and 21 females with an age range of 20–73 years (36.2 ± 19.3 years). Evaluation of DCJM model target detection efficacy The detection precision, recall, and F1-Score of target detection of drug-resistant LNTB detection in GUS images were 0.994. mAP_0.5 and mAP_0.5:0.95 of the GUS image-based DCJM model in drug-resistant LNTB detection were 0.995 and 0.897, respectively (Fig. 2 ). The ROC curves of the DCJM model built based on GUS images for the prediction of drug-resistant LNTB (including the training set, validation set, Test A and Test B) (Fig. 3 ). AUC = 0.992 (95% CI, 0.972-1.000) for the training set, AUC = 0.851 (95% CI, 0.733–0.948) for the validation set, AUC = 0.727 (95% CI, 0.488–0.924) for Test A, and AUC = 0.777 (95% CI, 0.644- 0.900) for Test B. The specific data on AUC, accuracy, sensitivity, specificity, PPV and NPV of the training set, validation set, Test A and Test B of the DCJM model for predicting drug-resistant LNTB are detailed in (Table 1 ). Table 1 DCJM model image classification diagnostic efficacy Groups AUC (95% CI) Accuracy Sensitivity Specificity PPV NPV Training set 0.992(0.972-1.000) 0.981 0.983 0.979 0.967 0.989 Validation set 0.851(0.733–0.948) 0.851 0.867 0.818 0.907 0.750 Test A 0.727(0.488–0.924) 0.700 0.857 0.652 0.429 0.938 Test B 0.777(0.644-0.900) 0.705 0.806 0.600 0.676 0.750 DCJM model image visualization As shown in Fig. 4 , six representative cases were selected in this study to demonstrate the visualization capabilities of the DCJM model. The DCJM model can automatically locate the nodes in the GUS image and depict the extent of the nodes using mask. With the heat map drawn by Grad-CAM, it can be observed that the focus of the DCJM model is on the lymph nodes and their marginal regions, and the red areas of the heat map indicate the areas of higher focus and the blue areas indicate the areas of lower focus. Discussion Drug-resistant TB patients have a long treatment cycle, low cure rate, and difficulty in control, which has posed a serious threat to global TB prevention and control 15 . In the available literature, fewer studies of extrapulmonary drug-resistant TB have been reported. A Korean study reported an overall resistance rate of extrapulmonary TB to at least one antituberculosis drug of 8.9% and a multidrug resistance rate of 1.8%, which is similar to or slightly lower than that of TB patients as a whole 5 . Currently, drug resistance detection of LNTB relies on drug susceptibility testing after MTB culture of biopsy tissues, which allows clinicians to understand the degree of sensitivity or resistance of MTB infected by patients to various anti-tuberculosis drugs, and is crucial for the diagnosis of drug-resistant tuberculosis. Existing MTB drug susceptibility testing techniques are mainly divided into phenotypic detection methods and molecular detection methods 16 . MTB phenotypic drug susceptibility technology is established on the basis of bacterial culture, and drug resistance is detected by observing the growth of MTB in drug-containing medium, but the long MTB culture time and low culture positivity rate have brought about difficulties in clinical diagnosis and treatment 17 . In the past 10 years, the development of MTB molecular drug sensitivity technology has been rapid. MTB molecular drug sensitivity technology is based on nucleic acid detection, the use of molecular biology technology to detect whether the relevant drug resistance genes of MTB are mutated, and the molecular detection technology represented by Gene Xpert MTB/RIF has high sensitivity and specificity, which can be used to diagnose MTB and detect drug resistance, but affected by the price, which restricts its use in primary hospitals 18 . influence, which restricts its development in primary hospitals 19 – 21 . Meanwhile, obtaining lymph node biopsy tissue specimens is an invasive method, and puncture injury may trigger the spread of MTB or even the formation of sinus tracts. Therefore, whether a non-invasive and accurate assessment method can be applied for LNTB drug resistance diagnosis is an urgent clinical problem to be solved. There are emerging AI algorithms that have been used for drug-resistant TB diagnosis in recent years. Most of the previous studies focused on pulmonary tuberculosis (TB), Liang 22 et al. who proposed DeepTB, a deep-learning based drug-resistant prediction system for TB, DeepTB achieved an AUC of 0.943 for drug-resistant TB diagnosis with a high degree of accuracy. In addition to medical images, genetic information can be used as a diagnostic tool for tuberculosis. Various molecular biology diagnostic techniques are capable of detecting drug resistance, which has been theorized to be caused by chromosomal mutations in TB, which are transmitted through genes. Also, rapid molecular tests using genomic information are more efficient than culture, so they are widely adopted and the relevant genetic data can be used for scientific research 23 . Most of the models that achieved better results were trained by gene sequences of MTB isolates 24 , but models built by gene sequence-based models are expensive and require high equipment. Deep learning using genomic data has been applied to reveal antibiotic resistance 25 , 26 . However, previous deep learning studies related in drug-resistant LNTB are less common. Therefore, in this study, the first GUS image-based DCJM model is proposed for prediction of drug-resistant LNTB. The model was used for prediction of drug-resistant LNTB by deep learning and mining GUS images of sensitive LNTB as well as drug-resistant LNTB. The DCJM model had a good performance in target detection with the detection precision, recall, F1-Score, mAP_0.5, and mAP_0.5. 0.95 were 0.994, 0.994, 0.994, 0.995, and 0.897, respectively, which showed that the network was able to accurately identify the target region in lymph node localization. Despite the large magnitude of variation in lesion regions and the large differences between different data, the network was still able to learn rich semantic information for accurate localization. The results of the DCJM model in predicting drug-resistant LNTB classification were as follows: the AUC = 0.992 (95% CI, 0.972-1.000) for the training set, and the AUC = 0.851 (95% CI, 0.733–0.948) for the training set, AUC = 0.727 (95% CI, 0.488–0.924) for Test A, and AUC = 0.777 (95% CI, 0.644-0.900) for Test B. Thus, the DCJM model has great potential as a simplified diagnostic for drug-resistant LNTB. It is envisaged that the model will be embedded in software or an online web server to enable smooth connection with the process of storing images in ultrasound. After inputting the ultrasound image, the algorithm is able to pinpoint lymph node lesions and identify LNTB resistance. With the help of the DCJM model, the radiologist can use the model output as a reference for image interpretation, and if there is a disagreement between the radiologist's diagnosis and the result given by the model, the radiologist can decide whether to change the result after reassessment. Prediction of drug-resistant TB by the DCJM model can help in the development of an individual treatment regimen based on the varying drug-resistance profiles. With the incidence of drug-resistant TB gradually increasing, therefore drug-resistant TB remains a critical issue that needs to be urgently addressed globally. The development of a reliable AI system using a large amount of ultrasound image data, which is more convenient than methods such as imaging histology as well as genetic information, can rapidly identify patients with drug-resistant TB and help clinicians make corresponding clinical decisions. However, the wide diffusion of deep learning models in hospitals is yet to be observed and validated with more samples. In this study, the AUC value of the external test set decreased compared to the internal validation set, so the DCJM model needs to be further trained to diagnose different models of ultrasound images as well as different types of patients. For the purpose of guiding clinical practice, the utility of the DCJM model needs to be tested in a real healthcare setting and further tightly integrated with routine ultrasound workflow, especially in poorer areas with a high burden of tuberculosis and a lack of advanced medical equipment and relevant specialized doctors. In the coming decades, better integration of AI with clinical workflows will have a significant impact on the whole process of TB from screening, diagnosis, treatment to prognosis, while saving healthcare resources, avoiding inappropriate treatments, and improving patients' quality of life. The limitations of this study are (1) the sample size is small and more cases need to be accumulated in subsequent studies to optimize the model, and the clinical diagnostic validity of the DCJM model will be tested in future clinical trials with more features; (2) this study only focused on GUS, whereas it is desirable for the DCJM model to extract more multimodal ultrasound images from multimodal ultrasound images including CDFI, elastic ultrasound and CEUS data; (3) the composite prediction model was not constructed by combining clinical indicators. In subsequent studies, we will collect more data from multimodal images and construct joint prediction models to improve the accuracy of LNTB resistance in the neck. Conclusion In summary, we have developed a GUS image-based DCJM that proceeds through a dual path of target detection's and image classification, and the model has been extensively evaluated in both internal and independent external test sets. Through continuous optimization and calibration, the model has great potential to be used as a powerful aid to facilitate the development of personalized treatment plans, leading to greater clinical benefit for patients. Declarations Author contributions CPJ contributed to the design of this study. LJL and YXY performed the study and collected data. ZWZ and YGY are responsible for reviewing the manuscript. All authors read and approved the final manuscript. Funding Source This research was funded by the Natural Science Foundation of Zhejiang Province (LTGY23H180005), the Major Project of Hangzhou Health Technology Plan (Z20230098), and the Hangzhou Biomedical and Health Industry Development Support Technology Special Project (2022WJC046). Competing interests The authors declare no competing interests. Ethics approval and consent to participate This study was in accordance with the Ethical Standards of the Institutional Ethics Committee of Hangzhou red cross hospital and with the 1964 Helsinki declaration and its later amendments or comparable Ethical Standards. All methods were carried out in accordance with relevant guidelines and regulations in the declaration. Written informed consent was provide by all included patients. Ethics batch number: [2023] Review No. (046). Availability of data and materials The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. References WHO. Global Tuberculosis Report 2023. (2023). Boonsarngsuk, V., Mangkang, K. & Santanirand, P. Prevalence and risk factors of drug-resistant extrapulmonary tuberculosis. Clin. Respir J. 12 , 2101–2109. 10.1111/crj.12779 (2018). Handa, U., Mundi, I. & Mohan, S. Nodal tuberculosis revisited: a review. J. Infect. Dev. Ctries. 6 , 6–12. 10.3855/jidc.2090 (2012). Cataño, J. C. & Robledo, J. Tuberculous Lymphadenitis and Parotitis. Microbiol. Spectr. 4 10.1128/microbiolspec.TNMI7-0008-2016 (2016). Lee, J. Y. Diagnosis and treatment of extrapulmonary tuberculosis. Tuberc Respir Dis. (Seoul) . 78 , 47–55. 10.4046/trd.2015.78.2.47 (2015). Chen, F. et al. Do as Sonographers Think: Contrast-enhanced Ultrasound for Thyroid Nodules Diagnosis via Microvascular Infiltrative Awareness. IEEE Trans. Med. Imaging Pp . 10.1109/tmi.2024.3405621 (2024). Shamir, S. B., Sasson, A. L., Margolies, L. R. & Mendelson, D. S. New Frontiers in Breast Cancer Imaging: The Rise of AI. Bioeng. (Basel) . 11 10.3390/bioengineering11050451 (2024). Liu, Z. et al. A deep learning model with data integration of ultrasound contrast-enhanced micro-flow cines, B-mode images, and clinical parameters for diagnosing significant liver fibrosis in patients with chronic hepatitis B. Eur. Radiol. 33 , 5871–5881. 10.1007/s00330-023-09436-z (2023). Chang, L. et al. An integrated nomogram combining deep learning, clinical characteristics and ultrasound features for predicting central lymph node metastasis in papillary thyroid cancer: A multicenter study. Front. Endocrinol. (Lausanne) . 14 , 964074. 10.3389/fendo.2023.964074 (2023). Sun, Q. et al. Deep Learning vs. Radiomics for Predicting Axillary Lymph Node Metastasis of Breast Cancer Using Ultrasound Images: Don't Forget the Peritumoral Region. Front. Oncol. 10 , 53. 10.3389/fonc.2020.00053 (2020). Lee, J. H. et al. Deep Learning-Based Computer-Aided Diagnosis System for Localization and Diagnosis of Metastatic Lymph Nodes on Ultrasound: A Pilot Study. Thyroid . 28 , 1332–1338. 10.1089/thy.2018.0082 (2018). Laverde-Saad, A. et al. Discriminative deep learning based benignity/malignancy diagnosis of dermatologic ultrasound skin lesions with pretrained artificial intelligence architecture. Skin. Res. Technol. 28 , 35–39. 10.1111/srt.13086 (2022). Farhat, M. et al. Drug-resistant tuberculosis: a persistent global health concern. Nat. Rev. Microbiol. 10.1038/s41579-024-01025-1 (2024). Zhou, B., Khosla, A., Lapedriza, A., Oliva, A. & Torralba, A. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2921–2929. Ye, M. et al. Antibiotic heteroresistance in Mycobacterium tuberculosis isolates: a systematic review and meta-analysis. Ann. Clin. Microbiol. Antimicrob. 20 , 73. 10.1186/s12941-021-00478-z (2021). Sanchini, A., Lanni, A., Giannoni, F. & Mustazzolu, A. Exploring diagnostic methods for drug-resistant tuberculosis: A comprehensive overview. Tuberculosis (Edinb) . 148 , 102522. 10.1016/j.tube.2024.102522 (2024). Mugumbate, G., Nyathi, B., Zindoga, A. & Munyuki, G. Application of Computational Methods in Understanding Mutations in Mycobacterium tuberculosis Drug Resistance. Front. Mol. Biosci. 8 , 643849. 10.3389/fmolb.2021.643849 (2021). Carvalho, A. C. et al. Differential diagnosis of cervical mycobacterial lymphadenitis in children. Pediatr. Infect. Dis. J. 29 , 629–633. 10.1097/INF.0b013e3181d1fdcd (2010). Osei Sekyere, J., Maphalala, N., Malinga, L. A., Mbelle, N. M. & Maningi, N. E. A Comparative Evaluation of the New Genexpert MTB/RIF Ultra and other Rapid Diagnostic Assays for Detecting Tuberculosis in Pulmonary and Extra Pulmonary Specimens. Sci. Rep. 9 , 16587. 10.1038/s41598-019-53086-5 (2019). Hao, X. et al. Cost-effectiveness analysis of Xpert in detecting Mycobacterium tuberculosis: A systematic review. Int. J. Infect. Dis. 95 , 98–105. 10.1016/j.ijid.2020.03.078 (2020). Silva, S. et al. Cost-effectiveness of Xpert®MTB/RIF in the diagnosis of tuberculosis: pragmatic study. Rev. Soc. Bras. Med. Trop. 54 , e07552020. 10.1590/0037-8682-0755-2020 (2021). Liang, S. et al. Deep learning for precise diagnosis and subtype triage of drug-resistant tuberculosis on chest computed tomography. MedComm (2020) . 5 , e487. 10.1002/mco2.487 (2024). Cohen, K. A., Manson, A. L., Desjardins, C. A., Abeel, T. & Earl, A. M. Deciphering drug resistance in Mycobacterium tuberculosis using whole-genome sequencing: progress, promise, and challenges. Genome Med. 11 , 45. 10.1186/s13073-019-0660-8 (2019). Yang, Y. et al. An end-to-end heterogeneous graph attention network for Mycobacterium tuberculosis drug-resistance prediction. Brief. Bioinform . 22 10.1093/bib/bbab299 (2021). Yang, Y. et al. DeepAMR for predicting co-occurrent resistance of Mycobacterium tuberculosis. Bioinformatics . 35 , 3240–3249. 10.1093/bioinformatics/btz067 (2019). Deelder, W. et al. Machine Learning Predicts Accurately Mycobacterium tuberculosis Drug Resistance From Whole Genome Sequencing Data. Front. Genet. 10 , 922. 10.3389/fgene.2019.00922 (2019). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted 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-5358428","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":376294773,"identity":"67068fcf-f735-4c77-b79a-de753f0531f2","order_by":0,"name":"Peijun Chen","email":"","orcid":"","institution":"Chinese and Western Hospital of Zhejiang Province (Hangzhou Red Cross Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Peijun","middleName":"","lastName":"Chen","suffix":""},{"id":376294775,"identity":"78c74044-93ab-4204-b631-c54546fcebf4","order_by":1,"name":"Jialei Luo","email":"","orcid":"","institution":"Chinese and Western Hospital of Zhejiang Province (Hangzhou Red Cross Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Jialei","middleName":"","lastName":"Luo","suffix":""},{"id":376294776,"identity":"209e2d39-59c6-4e87-95ac-454e6703263f","order_by":2,"name":"Xinyi Yan","email":"","orcid":"","institution":"Hangzhou Normal University","correspondingAuthor":false,"prefix":"","firstName":"Xinyi","middleName":"","lastName":"Yan","suffix":""},{"id":376294778,"identity":"76022843-5cc8-40fc-afa5-cbaa0af3813a","order_by":3,"name":"Ying Zhang","email":"","orcid":"","institution":"Chinese and Western Hospital of Zhejiang Province (Hangzhou Red Cross Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Zhang","suffix":""},{"id":376294779,"identity":"6c60fafb-bfa0-44ae-91a8-363e905bcda9","order_by":4,"name":"Wenzhi Zhang","email":"","orcid":"","institution":"Chinese and Western Hospital of Zhejiang Province (Hangzhou Red Cross Hospital)","correspondingAuthor":false,"prefix":"","firstName":"Wenzhi","middleName":"","lastName":"Zhang","suffix":""},{"id":376294781,"identity":"26d135d6-4dd9-48b1-b15a-d192dfc27938","order_by":5,"name":"Gaoyi Yang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA20lEQVRIiWNgGAWjYDADNvbmAwc+/CBJC8+xxIMze0iyRiLH+DAHGxEKDY6fPfzyS41NYh/PmQ+HGXgY5PnFDhDQciYvzVrmWFpiG3vvhsMFFgyGM2cnENByIMfMWILtcGIbz9kNh2fwMCQY3Cak5fwboJZ/QC0SOQ8O87ARo+VGjvHDj21gLQzEaZG88caMmbEvzbiN55gBMJAlCPuF73yO8ccf32xk57c3P/7w4YeNPL80AS0KBxjYpHkYGBwbIHwJ/MpBQL6BgfkjMJnYE1Y6CkbBKBgFIxYAADo0TVe7yceqAAAAAElFTkSuQmCC","orcid":"","institution":"Affiliated Hangzhou First People's Hospital, School of Medicine, Westlake University","correspondingAuthor":true,"prefix":"","firstName":"Gaoyi","middleName":"","lastName":"Yang","suffix":""}],"badges":[],"createdAt":"2024-10-30 05:23:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5358428/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5358428/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":70285176,"identity":"49eedc69-9014-4819-9667-cfedb06438c6","added_by":"auto","created_at":"2024-12-01 16:13:26","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":234881,"visible":true,"origin":"","legend":"\u003cp\u003eDiagrammatic representation of the DCJM model disease diagnosis method based on the construction of GUS images of lymph nodes\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5358428/v1/2eab7ce075d5ae1a1e87f8e7.jpg"},{"id":70285177,"identity":"ba60399a-1a45-43d5-95f5-9a8e3de9c1a8","added_by":"auto","created_at":"2024-12-01 16:13:26","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":291457,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic diagram of lymph node localization detection in GUS images\u003c/p\u003e\n\u003cp\u003eEvaluation of classification prediction efficacy of DCJM model\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5358428/v1/0fb59860e1a4ff7c4fb27f77.jpg"},{"id":70285175,"identity":"bafa6beb-c4a9-4839-bed6-3636ee0a7473","added_by":"auto","created_at":"2024-12-01 16:13:26","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":57961,"visible":true,"origin":"","legend":"\u003cp\u003eROC plot of grey scale ultrasound based DCJM model for prediction of resistant LNTB\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5358428/v1/17321a3518d5d3ff0d36bba7.jpg"},{"id":70285178,"identity":"fb22ff51-b91f-4358-987b-395e0c6991f8","added_by":"auto","created_at":"2024-12-01 16:13:26","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":266640,"visible":true,"origin":"","legend":"\u003cp\u003eDCJM model visualization heat map\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5358428/v1/96b8e46bc83b6737557e29bc.jpg"},{"id":70437262,"identity":"1e9213a1-5026-4d99-9a8e-885dd93b098e","added_by":"auto","created_at":"2024-12-03 07:17:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1217531,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5358428/v1/aeb82f3d-b7e6-4be8-8cac-1e74a60b3236.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Using the DCJM deep-learning model to diagnosis drug-resistant lymph node tuberculosis based on ultrasound images: A Multicenter Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDrug-resistant tuberculosis (TB) remains a threat to public health, with 410,000 (3.9%) multidrug/rifampicin-resistant TB patients globally in 2022, accounting for approximately 3.3% of new TB patients and 17% of retreatment patients\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. The estimated number of multidrug/rifampicin-resistant TB patients in China in 2022 is 30,000, and the rate of extrapulmonary TB resistance is higher than that of single pulmonary TB, making diagnostic and treatment evaluations difficult, and patients take longer (usually more than two years) and cost more to treat\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. One of the main challenges in controlling multidrug-resistant TB lies in the difficulty of diagnosing drug resistance in patients suspected of having TB, especially in first-time patients. Therefore, the diagnosis and treatment of drug-resistant TB is an urgent challenge in TB prevention and control.\u003c/p\u003e \u003cp\u003eLymph node tuberculosis (LNTB), as a common form of extrapulmonary TB, accounts for 35% of all extrapulmonary TB cases\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. LNTB is still treated with systemic anti-tuberculosis drug chemotherapy, which usually involves oral administration of multiple anti-tuberculosis drugs for up to 6\u0026ndash;9 months\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. With the proportion of drug-resistant TB increasing year by year, inadequate detection of drug-resistant TB in pathogenetically positive patients still exists, making subsequent treatment difficult. Early assessment of drug resistance in patients is a prerequisite for timely adjustment of treatment regimens. The main method of obtaining specimens for assessing LNTB drug resistance detection is ultrasound-guided lymph node biopsy, but it is invasive as well as at risk of various potential complications\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Therefore, it is urgent to explore a non-invasive screening method that can help improve the detection rate of drug-resistant LNTB.\u003c/p\u003e \u003cp\u003eIn the past few years, deep learning methods have made significant progress in the field of medical image analysis. Deep learning can significantly improve the recognition and prediction of medical images. It has achieved good performance in diagnosing thyroid, breast and liver diseases\u003csup\u003e\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. However, the application of deep learning in lymph node imaging is still relatively rare compared to other fields. Currently, a few studies have reported that deep learning based on ultrasound images of breast and thyroid nodes has been used to predict whether the associated draining lymph nodes are metastatic or not\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Lee et al.\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e collected 812 ultrasound images of cervical lymph nodes that had been pathologically confirmed to develop a computerized system with a view to help improve the accuracy of the diagnosis of lymph node metastasis, with an accuracy of 83%. In lymph node research, artificial intelligence (AI) has been compared to diagnostically experienced physicians and found to have similar diagnostic accuracy with adequate training\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThere are fewer reports of imaging studies on drug-resistant LNTB, and in this study, we attempted to construct a prediction model for drug-resistant LNTB in the neck by using a deep learning method based on grey-scale ultrasound (GUS) images to explore the value of the model in predicting drug-resistant LNTB.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003eEthics statement\u003c/p\u003e \u003cp\u003e The study was approved by the Medical Ethics Committee of Hangzhou Red Cross Hospital, and written informed consent was obtained from each participant (ApprovaI NO.: [2023] Review No. (046)).\u003c/p\u003e \u003cp\u003ePatient enrollment\u003c/p\u003e \u003cp\u003eIndependent cohort data from three hospitals in China were used in this study. LNTB GUS images from Hangzhou Red Cross Hospital between January 2018 and December 2022 were first retrospectively collected for inclusion in the study, and were randomly assigned proportionally to the training set (70%) and validation set (30%). Lymph node ultrasound images from Infectious Disease Hospital of Heilongjiang Province (Test A) and Kunming Third People's Hospital (Test B) between August 2022 and December 2022 were also collected as the external test set. Baseline characteristics of participants, including age, gender, lymph node status, and histological type. All data were anonymized to protect participant privacy. Patient inclusion criteria were as follows (1) tissue or pus culture and drug susceptibility testing techniques (liquid drug susceptibility testing), or molecular biology rapid resistance testing techniques (including linear probe, Gene-Xpert MTB/RIF assay, and other techniques) were used as diagnostic criteria for LNTB resistance\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. (2) The general information of the patients was complete. (3) Clear GUS images with distinguishable lesions manually outlined. Exclusion criteria included (1) Unsatisfactory quality of ultrasound images. (2) Lymph nodes that showed rupture phenomenon or were too large to be shown completely. (3) Inability to identify the patient's anti-tuberculosis drug resistance or sensitivity results. (4) Previous history of anti-tuberculosis treatment.\u003c/p\u003e \u003cp\u003eGUS image and clinical data acquisition\u003c/p\u003e \u003cp\u003eAll patients were placed in supine or lateral position for ultrasound examination. Region-by-region scanning was performed to look for suspected enlarged lymph nodes, and the largest lymph nodes were selected for observation. Two radiologists with 5 years of working experience in lymph node ultrasonography were responsible for acquiring the image data, and the lymph node images were saved in DICOM format. The images of each patient were obtained from the PACS database in DICOM format, and were organized based on the case number to obtain the corresponding examination results of the patient. A sheet of GUS static images of the largest longitudinal section of the target lymph node of each patient was selected for subsequent deep learning analysis. The corresponding examination results of the patients were obtained from the clinical case system, including tissue or pus culture, liquid drug susceptibility testing, Gene-Xpert MTB/RIF and whole genome sequencing were used as diagnostic criteria for LNTB resistance.\u003c/p\u003e \u003cp\u003eDevelopment and validation of deep learning model\u003c/p\u003e \u003cp\u003eAs the lymph node GUS image data has different disease classifications, data corruption, abnormal data distribution, and missing key information of the data, the data needs to be pre-processed when analyzing the data. It included format conversion, data screening and cleaning, data effective part extraction, data normalization, and lymph node GUS image enhancement. In addition by using small sample expansion techniques, the training set was expanded by rotating, flipping, scaling and panning the images, and generative adversarial networks. The combined use of these techniques helps to solve the problem of limited samples of lymph node ultrasound images and improve the performance of deep learning models.\u003c/p\u003e \u003cp\u003eWe proposed a target detection-image classification joint learning method (DCJM) model that combined target detection and image classification (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Firstly, the target detection model (YOLOv5 model) was used to detect the key regions of the input ultrasound image, and the high confidence and intersection of union (IoU) were set as the judgement conditions to improve the accuracy of the key region identification, to exclude most of the interfering factors and the influence of the wrong region, and then to classify the extracted key regions. At the same time, in order to avoid the problem that the key regions were not identified due to the high threshold conditions, and thus cannot participate in the classification without predicting the results, the DCJM model used the second image classification model to make auxiliary judgments to optimize the final overall classification results of the joint two-way. Adopting two-way detection, dynamically adjusting the allocation of prediction weights according to the final diagnostic efficacy results of the dataset, and combining the two-way prediction results for judgement, provides a new solution for solving the problem of misdiagnosis and omission of disease diagnosis.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo better explain the prediction logic of the YOLOv5 model, we used the gradient-weighted class activation mapping (Grad-CAM) method to visualize the most indicative image regions of the cervical lymph nodes by filtering the feature heatmap from the last convolutional layer\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Additionally, we used the t-distributed stochastic neighbor embedding (t-SNE) method to illustrate the overall prediction effect by converting the representation of the last layer of the deep neural network before the prediction node of each image in the test dataset into a shifted or non-shifted colour coding.\u003c/p\u003e \u003cp\u003eExperimental environment configuration:\u003c/p\u003e \u003cp\u003eHardware configuration: 12th Gen Intel(R), Core(TM), i9-12900KF (CPU), NVIDIA GeForce RTX 3090 (GPU), 64G (Memory).\u003c/p\u003e \u003cp\u003eSoftware configuration: Ubuntu 22.04, Python 3.9, Pytorch 1.12.0, CUDA 11.3.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003ePrecision, recall, F1 score, and, mean average precision (mAP) were used in the DCJM model to assess the performance of lymph node detection. IoU is used to measure the degree of overlap between the \"predicted border\" and the \"actual border\". mAP 0.5 is the average AP of all the images under each category when the IoU is 0.5. mAP 0.5:0.95 is the average AP of all the images under each category when the mAP threshold is from 50\u0026ndash;95%, and then average these ten values. 50% to a mAP threshold of 95%, with 5% intervals, 10 mAP values were obtained, and then these ten values were averaged. Area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were used to assess the predictive ability of the DCJM model for drug-resistant LNTB. The statistical analysis was performed using SPSS software version 25.0. The continuous variables in this study were presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD. All statistical significance levels were set as two-sided with p\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003ePatient characteristics\u003c/p\u003e \u003cp\u003eThe model training and internal validation set dataset was obtained from the Hangzhou Red Cross Hospital, and among the 92 patients with cervical LNTB in the drug-resistant group, there were 34 males and 58 females, with an age range of 19\u0026ndash;79 years (33.8\u0026thinsp;\u0026plusmn;\u0026thinsp;15.1 years); among the 142 patients with cervical LNTB in the sensitive group, there were 61 males and 81 females, with an age range of 17\u0026ndash;87 years (41.2\u0026thinsp;\u0026plusmn;\u0026thinsp;18.3 years). There were no significant differences between the two groups of patients in terms of age and gender (all \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.05). External test cohort A (Test A) consisted of 30 patients, which contained 7 resistant and 23 sensitive cases. There were 14 males and 16 females with an age range of 22\u0026ndash;72 years (35.9\u0026thinsp;\u0026plusmn;\u0026thinsp;13.1 years). External test cohort B (Test B) consisted of 33 patients, which contained 9 resistant and 24 sensitive cases. There were 12 males and 21 females with an age range of 20\u0026ndash;73 years (36.2\u0026thinsp;\u0026plusmn;\u0026thinsp;19.3 years).\u003c/p\u003e \u003cp\u003eEvaluation of DCJM model target detection efficacy\u003c/p\u003e \u003cp\u003eThe detection precision, recall, and F1-Score of target detection of drug-resistant LNTB detection in GUS images were 0.994. mAP_0.5 and mAP_0.5:0.95 of the GUS image-based DCJM model in drug-resistant LNTB detection were 0.995 and 0.897, respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe ROC curves of the DCJM model built based on GUS images for the prediction of drug-resistant LNTB (including the training set, validation set, Test A and Test B) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). AUC\u0026thinsp;=\u0026thinsp;0.992 (95% CI, 0.972-1.000) for the training set, AUC\u0026thinsp;=\u0026thinsp;0.851 (95% CI, 0.733\u0026ndash;0.948) for the validation set, AUC\u0026thinsp;=\u0026thinsp;0.727 (95% CI, 0.488\u0026ndash;0.924) for Test A, and AUC\u0026thinsp;=\u0026thinsp;0.777 (95% CI, 0.644- 0.900) for Test B. The specific data on AUC, accuracy, sensitivity, specificity, PPV and NPV of the training set, validation set, Test A and Test B of the DCJM model for predicting drug-resistant LNTB are detailed in (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \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\u003eDCJM model image classification diagnostic efficacy\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\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=\"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 \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroups\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC (95% CI)\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\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraining set\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.992(0.972-1.000)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.981\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.983\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.979\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.967\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.989\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation set\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.851(0.733\u0026ndash;0.948)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.851\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.867\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.818\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.907\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.750\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTest A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.727(0.488\u0026ndash;0.924)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.700\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.857\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.652\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.429\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.938\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTest B\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.777(0.644-0.900)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.705\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.806\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.600\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.676\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.750\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\u003eDCJM model image visualization\u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, six representative cases were selected in this study to demonstrate the visualization capabilities of the DCJM model. The DCJM model can automatically locate the nodes in the GUS image and depict the extent of the nodes using mask. With the heat map drawn by Grad-CAM, it can be observed that the focus of the DCJM model is on the lymph nodes and their marginal regions, and the red areas of the heat map indicate the areas of higher focus and the blue areas indicate the areas of lower focus.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eDrug-resistant TB patients have a long treatment cycle, low cure rate, and difficulty in control, which has posed a serious threat to global TB prevention and control\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. In the available literature, fewer studies of extrapulmonary drug-resistant TB have been reported. A Korean study reported an overall resistance rate of extrapulmonary TB to at least one antituberculosis drug of 8.9% and a multidrug resistance rate of 1.8%, which is similar to or slightly lower than that of TB patients as a whole\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eCurrently, drug resistance detection of LNTB relies on drug susceptibility testing after MTB culture of biopsy tissues, which allows clinicians to understand the degree of sensitivity or resistance of MTB infected by patients to various anti-tuberculosis drugs, and is crucial for the diagnosis of drug-resistant tuberculosis. Existing MTB drug susceptibility testing techniques are mainly divided into phenotypic detection methods and molecular detection methods\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. MTB phenotypic drug susceptibility technology is established on the basis of bacterial culture, and drug resistance is detected by observing the growth of MTB in drug-containing medium, but the long MTB culture time and low culture positivity rate have brought about difficulties in clinical diagnosis and treatment\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. In the past 10 years, the development of MTB molecular drug sensitivity technology has been rapid. MTB molecular drug sensitivity technology is based on nucleic acid detection, the use of molecular biology technology to detect whether the relevant drug resistance genes of MTB are mutated, and the molecular detection technology represented by Gene Xpert MTB/RIF has high sensitivity and specificity, which can be used to diagnose MTB and detect drug resistance, but affected by the price, which restricts its use in primary hospitals\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. influence, which restricts its development in primary hospitals\u003csup\u003e\u003cspan additionalcitationids=\"CR20\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Meanwhile, obtaining lymph node biopsy tissue specimens is an invasive method, and puncture injury may trigger the spread of MTB or even the formation of sinus tracts. Therefore, whether a non-invasive and accurate assessment method can be applied for LNTB drug resistance diagnosis is an urgent clinical problem to be solved.\u003c/p\u003e \u003cp\u003eThere are emerging AI algorithms that have been used for drug-resistant TB diagnosis in recent years. Most of the previous studies focused on pulmonary tuberculosis (TB), Liang\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e et al. who proposed DeepTB, a deep-learning based drug-resistant prediction system for TB, DeepTB achieved an AUC of 0.943 for drug-resistant TB diagnosis with a high degree of accuracy. In addition to medical images, genetic information can be used as a diagnostic tool for tuberculosis. Various molecular biology diagnostic techniques are capable of detecting drug resistance, which has been theorized to be caused by chromosomal mutations in TB, which are transmitted through genes. Also, rapid molecular tests using genomic information are more efficient than culture, so they are widely adopted and the relevant genetic data can be used for scientific research\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Most of the models that achieved better results were trained by gene sequences of MTB isolates\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, but models built by gene sequence-based models are expensive and require high equipment. Deep learning using genomic data has been applied to reveal antibiotic resistance\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eHowever, previous deep learning studies related in drug-resistant LNTB are less common. Therefore, in this study, the first GUS image-based DCJM model is proposed for prediction of drug-resistant LNTB. The model was used for prediction of drug-resistant LNTB by deep learning and mining GUS images of sensitive LNTB as well as drug-resistant LNTB. The DCJM model had a good performance in target detection with the detection precision, recall, F1-Score, mAP_0.5, and mAP_0.5. 0.95 were 0.994, 0.994, 0.994, 0.995, and 0.897, respectively, which showed that the network was able to accurately identify the target region in lymph node localization. Despite the large magnitude of variation in lesion regions and the large differences between different data, the network was still able to learn rich semantic information for accurate localization. The results of the DCJM model in predicting drug-resistant LNTB classification were as follows: the AUC\u0026thinsp;=\u0026thinsp;0.992 (95% CI, 0.972-1.000) for the training set, and the AUC\u0026thinsp;=\u0026thinsp;0.851 (95% CI, 0.733\u0026ndash;0.948) for the training set, AUC\u0026thinsp;=\u0026thinsp;0.727 (95% CI, 0.488\u0026ndash;0.924) for Test A, and AUC\u0026thinsp;=\u0026thinsp;0.777 (95% CI, 0.644-0.900) for Test B. Thus, the DCJM model has great potential as a simplified diagnostic for drug-resistant LNTB. It is envisaged that the model will be embedded in software or an online web server to enable smooth connection with the process of storing images in ultrasound. After inputting the ultrasound image, the algorithm is able to pinpoint lymph node lesions and identify LNTB resistance. With the help of the DCJM model, the radiologist can use the model output as a reference for image interpretation, and if there is a disagreement between the radiologist's diagnosis and the result given by the model, the radiologist can decide whether to change the result after reassessment. Prediction of drug-resistant TB by the DCJM model can help in the development of an individual treatment regimen based on the varying drug-resistance profiles.\u003c/p\u003e \u003cp\u003eWith the incidence of drug-resistant TB gradually increasing, therefore drug-resistant TB remains a critical issue that needs to be urgently addressed globally. The development of a reliable AI system using a large amount of ultrasound image data, which is more convenient than methods such as imaging histology as well as genetic information, can rapidly identify patients with drug-resistant TB and help clinicians make corresponding clinical decisions. However, the wide diffusion of deep learning models in hospitals is yet to be observed and validated with more samples. In this study, the AUC value of the external test set decreased compared to the internal validation set, so the DCJM model needs to be further trained to diagnose different models of ultrasound images as well as different types of patients. For the purpose of guiding clinical practice, the utility of the DCJM model needs to be tested in a real healthcare setting and further tightly integrated with routine ultrasound workflow, especially in poorer areas with a high burden of tuberculosis and a lack of advanced medical equipment and relevant specialized doctors. In the coming decades, better integration of AI with clinical workflows will have a significant impact on the whole process of TB from screening, diagnosis, treatment to prognosis, while saving healthcare resources, avoiding inappropriate treatments, and improving patients' quality of life.\u003c/p\u003e \u003cp\u003eThe limitations of this study are (1) the sample size is small and more cases need to be accumulated in subsequent studies to optimize the model, and the clinical diagnostic validity of the DCJM model will be tested in future clinical trials with more features; (2) this study only focused on GUS, whereas it is desirable for the DCJM model to extract more multimodal ultrasound images from multimodal ultrasound images including CDFI, elastic ultrasound and CEUS data; (3) the composite prediction model was not constructed by combining clinical indicators. In subsequent studies, we will collect more data from multimodal images and construct joint prediction models to improve the accuracy of LNTB resistance in the neck.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn summary, we have developed a GUS image-based DCJM that proceeds through a dual path of target detection's and image classification, and the model has been extensively evaluated in both internal and independent external test sets. Through continuous optimization and calibration, the model has great potential to be used as a powerful aid to facilitate the development of personalized treatment plans, leading to greater clinical benefit for patients.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCPJ contributed to the design of this study. LJL and YXY performed the study and collected data. ZWZ and YGY are responsible for reviewing the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding Source\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was funded by the Natural Science Foundation of Zhejiang Province (LTGY23H180005), the Major Project of Hangzhou Health Technology Plan (Z20230098), and the Hangzhou Biomedical and Health Industry Development Support Technology Special Project (2022WJC046).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was in accordance with the Ethical Standards of the Institutional Ethics Committee of Hangzhou red cross hospital and with the 1964 Helsinki declaration and its later amendments or comparable Ethical Standards. All methods were carried out in accordance with relevant guidelines and regulations in the declaration. Written informed consent was provide by all included patients. Ethics batch number: [2023] Review No. (046).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWHO. Global Tuberculosis Report 2023. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoonsarngsuk, V., Mangkang, K. \u0026amp; Santanirand, P. Prevalence and risk factors of drug-resistant extrapulmonary tuberculosis. \u003cem\u003eClin. Respir J.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 2101\u0026ndash;2109. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/crj.12779\u003c/span\u003e\u003cspan address=\"10.1111/crj.12779\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHanda, U., Mundi, I. \u0026amp; Mohan, S. Nodal tuberculosis revisited: a review. \u003cem\u003eJ. Infect. Dev. Ctries.\u003c/em\u003e \u003cb\u003e6\u003c/b\u003e, 6\u0026ndash;12. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3855/jidc.2090\u003c/span\u003e\u003cspan address=\"10.3855/jidc.2090\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCata\u0026ntilde;o, J. C. \u0026amp; Robledo, J. Tuberculous Lymphadenitis and Parotitis. \u003cem\u003eMicrobiol. Spectr.\u003c/em\u003e \u003cb\u003e4\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1128/microbiolspec.TNMI7-0008-2016\u003c/span\u003e\u003cspan address=\"10.1128/microbiolspec.TNMI7-0008-2016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, J. Y. Diagnosis and treatment of extrapulmonary tuberculosis. \u003cem\u003eTuberc Respir Dis. (Seoul)\u003c/em\u003e. \u003cb\u003e78\u003c/b\u003e, 47\u0026ndash;55. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.4046/trd.2015.78.2.47\u003c/span\u003e\u003cspan address=\"10.4046/trd.2015.78.2.47\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, F. et al. Do as Sonographers Think: Contrast-enhanced Ultrasound for Thyroid Nodules Diagnosis via Microvascular Infiltrative Awareness. \u003cem\u003eIEEE Trans. Med. Imaging Pp\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/tmi.2024.3405621\u003c/span\u003e\u003cspan address=\"10.1109/tmi.2024.3405621\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShamir, S. B., Sasson, A. L., Margolies, L. R. \u0026amp; Mendelson, D. S. New Frontiers in Breast Cancer Imaging: The Rise of AI. \u003cem\u003eBioeng. (Basel)\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/bioengineering11050451\u003c/span\u003e\u003cspan address=\"10.3390/bioengineering11050451\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, Z. et al. A deep learning model with data integration of ultrasound contrast-enhanced micro-flow cines, B-mode images, and clinical parameters for diagnosing significant liver fibrosis in patients with chronic hepatitis B. \u003cem\u003eEur. Radiol.\u003c/em\u003e \u003cb\u003e33\u003c/b\u003e, 5871\u0026ndash;5881. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00330-023-09436-z\u003c/span\u003e\u003cspan address=\"10.1007/s00330-023-09436-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang, L. et al. An integrated nomogram combining deep learning, clinical characteristics and ultrasound features for predicting central lymph node metastasis in papillary thyroid cancer: A multicenter study. \u003cem\u003eFront. Endocrinol. (Lausanne)\u003c/em\u003e. \u003cb\u003e14\u003c/b\u003e, 964074. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fendo.2023.964074\u003c/span\u003e\u003cspan address=\"10.3389/fendo.2023.964074\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun, Q. et al. Deep Learning vs. Radiomics for Predicting Axillary Lymph Node Metastasis of Breast Cancer Using Ultrasound Images: Don't Forget the Peritumoral Region. \u003cem\u003eFront. Oncol.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, 53. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fonc.2020.00053\u003c/span\u003e\u003cspan address=\"10.3389/fonc.2020.00053\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, J. H. et al. Deep Learning-Based Computer-Aided Diagnosis System for Localization and Diagnosis of Metastatic Lymph Nodes on Ultrasound: A Pilot Study. \u003cem\u003eThyroid\u003c/em\u003e. \u003cb\u003e28\u003c/b\u003e, 1332\u0026ndash;1338. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1089/thy.2018.0082\u003c/span\u003e\u003cspan address=\"10.1089/thy.2018.0082\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLaverde-Saad, A. et al. Discriminative deep learning based benignity/malignancy diagnosis of dermatologic ultrasound skin lesions with pretrained artificial intelligence architecture. \u003cem\u003eSkin. Res. Technol.\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e, 35\u0026ndash;39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/srt.13086\u003c/span\u003e\u003cspan address=\"10.1111/srt.13086\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarhat, M. et al. Drug-resistant tuberculosis: a persistent global health concern. \u003cem\u003eNat. Rev. Microbiol.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41579-024-01025-1\u003c/span\u003e\u003cspan address=\"10.1038/s41579-024-01025-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou, B., Khosla, A., Lapedriza, A., Oliva, A. \u0026amp; Torralba, A. in \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition.\u003c/em\u003e 2921\u0026ndash;2929.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYe, M. et al. Antibiotic heteroresistance in Mycobacterium tuberculosis isolates: a systematic review and meta-analysis. \u003cem\u003eAnn. Clin. Microbiol. Antimicrob.\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e, 73. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12941-021-00478-z\u003c/span\u003e\u003cspan address=\"10.1186/s12941-021-00478-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSanchini, A., Lanni, A., Giannoni, F. \u0026amp; Mustazzolu, A. Exploring diagnostic methods for drug-resistant tuberculosis: A comprehensive overview. \u003cem\u003eTuberculosis (Edinb)\u003c/em\u003e. \u003cb\u003e148\u003c/b\u003e, 102522. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.tube.2024.102522\u003c/span\u003e\u003cspan address=\"10.1016/j.tube.2024.102522\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMugumbate, G., Nyathi, B., Zindoga, A. \u0026amp; Munyuki, G. Application of Computational Methods in Understanding Mutations in Mycobacterium tuberculosis Drug Resistance. \u003cem\u003eFront. Mol. Biosci.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 643849. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fmolb.2021.643849\u003c/span\u003e\u003cspan address=\"10.3389/fmolb.2021.643849\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarvalho, A. C. et al. Differential diagnosis of cervical mycobacterial lymphadenitis in children. \u003cem\u003ePediatr. Infect. Dis. J.\u003c/em\u003e \u003cb\u003e29\u003c/b\u003e, 629\u0026ndash;633. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/INF.0b013e3181d1fdcd\u003c/span\u003e\u003cspan address=\"10.1097/INF.0b013e3181d1fdcd\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOsei Sekyere, J., Maphalala, N., Malinga, L. A., Mbelle, N. M. \u0026amp; Maningi, N. E. A Comparative Evaluation of the New Genexpert MTB/RIF Ultra and other Rapid Diagnostic Assays for Detecting Tuberculosis in Pulmonary and Extra Pulmonary Specimens. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 16587. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-019-53086-5\u003c/span\u003e\u003cspan address=\"10.1038/s41598-019-53086-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHao, X. et al. Cost-effectiveness analysis of Xpert in detecting Mycobacterium tuberculosis: A systematic review. \u003cem\u003eInt. J. Infect. Dis.\u003c/em\u003e \u003cb\u003e95\u003c/b\u003e, 98\u0026ndash;105. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ijid.2020.03.078\u003c/span\u003e\u003cspan address=\"10.1016/j.ijid.2020.03.078\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSilva, S. et al. Cost-effectiveness of Xpert\u0026reg;MTB/RIF in the diagnosis of tuberculosis: pragmatic study. \u003cem\u003eRev. Soc. Bras. Med. Trop.\u003c/em\u003e \u003cb\u003e54\u003c/b\u003e, e07552020. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1590/0037-8682-0755-2020\u003c/span\u003e\u003cspan address=\"10.1590/0037-8682-0755-2020\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang, S. et al. Deep learning for precise diagnosis and subtype triage of drug-resistant tuberculosis on chest computed tomography. \u003cem\u003eMedComm (2020)\u003c/em\u003e. \u003cb\u003e5\u003c/b\u003e, e487. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/mco2.487\u003c/span\u003e\u003cspan address=\"10.1002/mco2.487\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohen, K. A., Manson, A. L., Desjardins, C. A., Abeel, T. \u0026amp; Earl, A. M. Deciphering drug resistance in Mycobacterium tuberculosis using whole-genome sequencing: progress, promise, and challenges. \u003cem\u003eGenome Med.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 45. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13073-019-0660-8\u003c/span\u003e\u003cspan address=\"10.1186/s13073-019-0660-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, Y. et al. An end-to-end heterogeneous graph attention network for Mycobacterium tuberculosis drug-resistance prediction. \u003cem\u003eBrief. Bioinform\u003c/em\u003e. \u003cb\u003e22\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bib/bbab299\u003c/span\u003e\u003cspan address=\"10.1093/bib/bbab299\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, Y. et al. DeepAMR for predicting co-occurrent resistance of Mycobacterium tuberculosis. \u003cem\u003eBioinformatics\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e, 3240\u0026ndash;3249. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bioinformatics/btz067\u003c/span\u003e\u003cspan address=\"10.1093/bioinformatics/btz067\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeelder, W. et al. Machine Learning Predicts Accurately Mycobacterium tuberculosis Drug Resistance From Whole Genome Sequencing Data. \u003cem\u003eFront. Genet.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, 922. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fgene.2019.00922\u003c/span\u003e\u003cspan address=\"10.3389/fgene.2019.00922\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Ultrasound, Deep learning, Lymph node tuberculosis, Drug resistance","lastPublishedDoi":"10.21203/rs.3.rs-5358428/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5358428/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eOur objective was to develop a deep learning model based on grey-scale ultrasound (GUS) images for predicting whether lymph node tuberculosis (LNTB) of the neck is drug resistant. The GUS images of 297 cases of cervical LNTB confirmed to be drug-resistant or sensitive by laboratory examination in three hospitals were retrospectively collected. A target detection-image classification joint learning method (DCJM) model combining target detection and image classification was constructed from the training set, and the diagnostic efficacy of the DCJM model was evaluated by the data from the internal validation set, Test A and Test B. We used mean average precision (mAP) to assess the accuracy of target detection in the DCJM model, The mAP_0.5 and mAP_0.5:0.95 of the DCJM model for LNTB detection were 0.995 and 0.897, respectively. The area under the curve (AUC) of this model in the training set, validation set, Test A, and Test B were 0.992 (95% CI, 0.972-1.000), 0.851 (95% CI, 0.733\u0026ndash;0.948), 0.727 (95% CI, 0.488\u0026ndash;0.924), and 0.777 (95% CI, 0.644-0.900), respectively. The DCJM model has a strong detection function as well as a good predictive value for drug-resistant LNTB, providing valuable information for individualized treatment decisions.\u003c/p\u003e","manuscriptTitle":"Using the DCJM deep-learning model to diagnosis drug-resistant lymph node tuberculosis based on ultrasound images: A Multicenter Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-01 16:13:20","doi":"10.21203/rs.3.rs-5358428/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"63c7321b-c4be-4a6a-b6d8-bced2b8295b3","owner":[],"postedDate":"December 1st, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":40056439,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":40056440,"name":"Health sciences/Diseases"}],"tags":[],"updatedAt":"2024-12-03T07:09:20+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-01 16:13:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5358428","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5358428","identity":"rs-5358428","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}