Classification of lumbar spondylosis from MRI images using CNN ensemble method

Classification of lumbar spondylosis from MRI images using CNN ensemble method | 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 Classification of lumbar spondylosis from MRI images using CNN ensemble method Ewunate Assaye Kassaw, Bekele Mulat Enyew, Abebe Alemu Abitew, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-2753236/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 Background : Due to an unfavorable ratio between the mechanical load and the size of the intervertebral discs, lumbar spondylosis, one of the most common causes of morbidity and disability. The preferred imaging technique for determining the origins of complex lower back pain is MRI. Healthcare systems in underdeveloped countries have a shortage of radiologists. Developing a CNN ensemble model for diagnosing lumbar spondylosis from MRI images was the aim of this study. Methods : 11158 T1 and T2 labeled MRI scans were collected from the University of Gondar specialized hospital and prepared for image processing. Since the median filter performed better than the others, it was chosen to denoise the data. The data was then augmented and split into an 80:20 train test ratio. An ensemble model was built by concatenating the proposed CNN and VGG19 models. Finally, the model was deployed. Results : An ensemble model achieved strong performance of 98.16% accuracy, 98% recall, and 98% precision. The GUI provides a setting appropriate for routine model usage. Conclusion : The research confirms that lumbar spondylosis can be diagnosed using MRI data and a CNN ensemble model. Physical sciences/Engineering/Biomedical engineering Physical sciences/Engineering/Electrical and electronic engineering Physical sciences/Engineering CNN ensemble technique image processing lumbar spondylosis MRI Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction The number of years with a disability caused by lower back pain increased by 54% between 1990 and 2015, with the highest rise occurring in third-world nations [ 1 ]. Between 50% and 80% of people are thought to have lower back ache in their lives caused by intervertebral disc degeneration and damage which imposes a huge socioeconomic burden on the neighborhood [ 1 ], [ 2 ]. The lumbar area is predominantly affected by spondylosis, a common disease of the spinal column, mostly due to an unfavorable ratio between the mechanical loads [ 2 ]. Lumbar spondylosis (LS) is becoming a significant clinical problem that the world is currently facing. Lumbar spondylosis symptoms include degeneration of the intervertebral discs or facet joints, formation of bony spurs (osteophytes), sclerosis of the vertebral body or endplates, hypertrophy of ligaments, or, in more severe cases, narrowing of the spinal line or disk space [ 3 ], [ 4 ]. A history, physical exam, and variety of imaging techniques are carried out to diagnose LS condition [ 5 ]. The basis for assessing lower back pain is currently MRI rather than CT and X-ray imaging because of the availability of soft tissue features, cross-section capacity, and the absence of hazardous radiation [ 5 ]. Several studies have been conducted using machine learning, especially CNNs, for the LS classification and diagnostic tasks, despite the fact that the majority of them used insufficient datasets and neglected important picture preprocessing techniques [ 6 ]–[ 12 ]. Columbo software is used for segmentation but more research is needed to see whether it can genuinely improve outcomes standardization, reduce the reading time, and increase inter-rater agreement [ 6 ]. According to a research team, the herniated lumbar disc can be classified as foraminal, post-lateral, or median using the VGG16 end-to-end classifier model. They did not include sagittal views, scans, or other disorders that were essential to the investigation, nor did they employ any extra pretrained models [ 8 ]. An improved model was developed using the TFLite model optimization technique. The model skips crucial image processing methods like noise removal that could improve the model's performance and only uses SoftMax as a classifier [ 9 ]. Two separate deep neural networks, AlexNet and GoogLeNet, were utilized to diagnose two classes of spondylolisthesis. The results show that GoogleLeNet outperforms AlexNet when trained with private X-ray images. Implementing crucial strategies like feature segmentation and classification techniques wasn't demonstrated [ 10 ]. Previous studies that simply examined sagittal T2-weighted images were unable to accurately detect spinal canal stenosis and modic alterations, necessitating the use of both T-1 and T-2 weighted scans. While analyzing lumbar spine MRIs in clinical settings, it is also necessary to take into account automatic disc herniation or nerve root compression identification [ 6 ], [ 8 ], [ 11 ]. Healthcare systems in underdeveloped countries have a shortage of radiologists [ 13 ]. Smaller data sets, the exclusion of fundamental image processing methods, and a failure to account for the use of ensemble CNN models for enhanced performance are some of the shortcomings in the machine learning models developed so far. Using an MRI dataset that had undergone considerable preprocessing, the aim of this study was to develop a deep learning model by merging CNN models to identify the various LS illness classifications. 2. Methods An ensemble model was developed by combining the CNN model with VGG19 to detect and diagnose Lumbar spondylosis from MRI image. To develop this ensemble model data were collected, the date was prepared, preprocessed, features extracted, and delivered to the training algorithm. The test data was then given into the classifier, whose performance was assessed using a variety of metrics, and the model was deployed using the Flask environment. Figure 2.1 shows the general block diagram of the method used in this study. 2.1. Dataset preparation The specialized hospital of the university of Gondar provided 11158 T1 and T2 weighted axial and sagittal sequence Lumbosacral MRI images that were scanned between October 27, 2021, and September 6, 2022. The data set contains 2739 normal, 2841 disc-bulge, 2804 disc-desiccation, and 2774 nerve root compression (NRC) images. Then the data were converted from DICOM to PNG format for the convenience of image processing [ 14 ]. The initial RGBA MRI image format is then converted to three channel format which is suitable for further processing. The required picture portions were crop using array slices, and extra annotations were minimized. To take the device's computing power into account, the 64x64 pixel resolution was then obtained. Finally, using label encoder, the data was labeled as LS0 = bulge, LS1 = desiccation, LS2 = normal, and LS3 = NRC. 2.2. Image preprocessing MRI image is susceptible to salt-and-pepper, gaussian, Rician and thermal noises which can be removed by keeping the quality and enhancing the smoothness of image using filters [ 15 ], [ 16 ]. For filters these noises gaussian, mean, median, mean-median, and adaptive histogram equalization methods are reported to performe better for MRI imges [ 6 ], [ 8 ], [ 10 ], [ 11 ]. Following the individual application of these filters, the overall system performance was assessed as shown in Table 2.1 , and the median filter outperformed the others. Table 2.1 The performance of the model after using different filtering methods Name Filtering techniques Result /accuracy in% Training Validation Testing AUC ROC Experiment result after filtering preprocess Gaussian filter 100 98.61 97.41 98.32 Mean filter 100 98.61 97.04 98.13 Median filter 100 97.69 97.78 98.59 Mean-median filter 99.65 97.22 95.19 96.72 AHE 99.88 95.37 96.30 97.55 Assumed to be the task after denoising, image segmentation was left out of this study because its implementation did not result in any improvements. To increase the training dataset and avoid overfitting rotation, shifting, adding brightness, flipping, and zooming offline data augmentation methods were used. 2.3. Model development The dataset was successfully split into an 80:20 ratio utilizing hold out approach after the MRI images were correctly preprocessed. There were 7810, 1674, and 1674 datasets distributed for training, validation, and testing, respectively. SoftMax classifier was used to assess a few pre-trained models at epochs of 100, and the result shows VGG net performed better with an accuracy value of 93.32%. Then, features are extracted from MRI images using the suggested CNN model and proposed VGG19 net, in parallel. Then, in order to create a model with improved performance, the developed CNN classifier model and the pretrained VGG19 model was combined. The classifier receives all of the features that were retrieved from the two models in one form. The hyperparameters are modified with an Adam optimizer, categorical cross-entropy, learning rate of 0.001, at epochs of 100, and a batch size of 32. Figure 2.2 shows the overview of the proposed model development. The accuracy, precision, recall, F-Measure, confusion matrices, specificity, and AUC ROC curve metrics were used to assess the model's performance. Finally, a graphical user interface was developed using a flask environment to classify an input image as belonging to one of the classes and offer a probability percentage out of 100%. 3. Experimental Result And Discussion 3.1. Results 3.1.1. Experiment without image preprocessing Table 3.1 shows the outcomes of tests conducted using proposed CNN, pretrained VGG19, and ensemble models without using image preprocessing. Table 3.1 Experiment result before data preprocess Name Metrics and model name Test accuracy AUC-ROC Experiment result before filtering, segmentation and augmentation Proposed CNN model 91.85% 94.20% Proposed VGG19 model 90.37% 93.31% Proposed ensemble model 94.44% 96.10% 3.1.2. Experiment result for various parameters A comprehensive experiment has been carried out a to determine the best parameters. Figure 3 .2 shows an overview of comparisons for a few parameters. 3.1.3. Final model detail experimental result discussion The proposed ensemble model was found to be the most ideal model at 100 epochs for classification of lumbar spondylosis following augmentation as shown in Table 3.2 . Table 3.2 Experimental result for final model Name Model name Result % Training validation testing Experiment result for final ensemble model Proposed ensemble 100 98.32 98.16 Figure 3.3 shows the results of testing the proposed ensemble method using the preprocessed image. The precision, recall and F1-score results are also shown in Table 3.3 . Table 3.3 Precision, Recall and F1-score results of the proposed ensemble method Precision Recall F1-Score Bulge 0.99 0.99 0.99 Desiccation 0.98 0.97 0.98 Normal 0.97 0.98 0.98 NRC 0.99 0.99 0.99 Accuracy 0.98 3.1.4. Model Deployment The deployment overview of the model user interface is shown in Fig. 3.4. 3.2. Discussion In this study ensemble method developed by concatenating proposed CNN and VGG19 was used to classify lumbar spondylosis. Initially the performance of the proposed CNN, VGG19 and ensemble models were tested and scored an accuracy result of 91.85%, 90.37%, and 94.44% respectively. The accuracy vs. epoch and loss vs. epoch plots in Fig. 3.1 shows that accuracy increases and loss decreases as epoch increases and epoch size 100 is an ideal for the proposed model. Moreover, the proposed ensemble model accuracy was tested and resulted better result for train test split of 80/20, batch size of 32, learning rate of 0.001, dropout of 0.1, and max-pooling mechanism as shown in Fig. 3.1 . To improve the performance further image preprocessing and augmentation processes were done on the input MRI data. When tested using the confusion matrix provided in Fig. 3.3 , the proposed model achieved a weighted average value of 98.16%, 98%, 98%, and 98% for accuracy, precision recall, and F1-score, respectively. This outcome demonstrates how preprocessing and augmentation techniques significantly increased the model's accuracy. The model accuracy loss graphs and the AUC ROC curve also demonstrate how preprocessing and augmentation considerably enhanced the performance of the proposed model. On test data, the AUC ROC score was 98.78%, indicating the model can classify between class labels accurately. A convenient environment for real-time use was created by the deployment of the built model. 4. Conclusion One of the issues the globe is facing is lumbar spondylosis (LS), which is becoming into a severe clinical concern. Due to the availability of soft tissue information, cross-sectional capacity, and lack of hazardous radiation, MRI is now the primary method used to study LS. The row MRI image was augmented and preprocessed, and the proposed CNN and VGG19 models concatenated to develop an ensemble model. Testing the ensemble model's accuracy resulted in a 98.16% accuracy score, which is a promising result for using it to diagnose lumbar spondylosis. Declarations Ethics approval and consent to participate Ethical approval letters were obtained from Institutional review boards of University of Gondar institute of health science. All methods were carried out in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Consent for publication This research did not involve humans, animals, or other subjects Availability of data and material The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request Competing interests The authors declare no competing interests. Funding Not applicable Authors' contributions BM and EA conceptualized, designed, and implemented in collaboration with the co-investigators AA and YG. All authors contributed to the preliminary study, the design, prototyping, and testing. The article was drafted by BM and EA, taking into account the comments and suggestions of the co-authors. All co-authors had the opportunity to comment on the manuscript and approved the final version for publication. Acknowledgments The specialized hospital at the University of Gondar provided the MRI data sets needed to complete the investigation. Much assistance was provided by the Department of Information Technology at University of Gondar. References M. Kahere and T. Ginindza, “Mapping evidence on the prevalence, incidence, risk factors and cost associated with chronic low back pain among adults in Sub-Saharan Africa: a systematic scoping review protocol,” Syst. Rev., vol. 9, no. 1, p. 57, Mar. 2020, doi: 10.1186/s13643-020-01321-w . M. Kolenkiewicz, A. Włodarczyk, and J. Wojtkiewicz, “Diagnosis and Incidence of Spondylosis and Cervical Disc Disorders in the University Clinical Hospital in Olsztyn, in Years 2011–2015,” BioMed Res. Int. , vol. 2018, p. 5643839, Mar. 2018, doi: 10.1155/2018/5643839 . R. Tsujimoto et al. , “Prevalence of lumbar spondylosis and its association with low back pain among community-dwelling Japanese women,” BMC Musculoskelet. Disord., vol. 17, no. 1, p. 493, Dec. 2016, doi: 10.1186/s12891-016-1343-x . C. Barlotta, “Prevalence and Risk Factors for Lumbar Spondylosis and Its Association with Low Back Pain among Rural Korean Residents.,” Minerva Stomatol. , vol. 15, no. 6, pp. 458–461, Jun. 1966. M. Munir Mir, A. John, M. Naeem, H. Butt, and A. Ali, “Prevalence and Radiological Evaluation of Lumbar Spondylosis on Magnetic Resonance Imaging,” EAS J. Radiol. Imaging Technol. , vol. 3, pp. 57–65, Apr. 2021, doi: 10.36349/easjrit.2021.v03i02.005 . N. C. Lehnen et al. , “Detection of Degenerative Changes on MR Images of the Lumbar Spine with a Convolutional Neural Network: A Feasibility Study,” Diagnostics , vol. 11, no. 5, Art. no. 5, May 2021, doi: 10.3390/diagnostics11050902 . A. Hirschmann, J. Cyriac, B. Stieltjes, T. Kober, J. Richiardi, and P. Omoumi, “Artificial Intelligence in Musculoskeletal Imaging: Review of Current Literature, Challenges, and Trends,” Semin. Musculoskelet. Radiol. , vol. 23, no. 03, pp. 304–311, Jun. 2019, doi: 10.1055/s-0039-1684024 . W. Mbarki, M. Bouchouicha, S. Frizzi, F. Tshibasu, L. B. Farhat, and M. Sayadi, “Lumbar spine discs classification based on deep convolutional neural networks using axial view MRI,” Interdiscip. Neurosurg., vol. 22, p. 100837, Dec. 2020, doi: 10.1016/j.inat.2020.100837 . D. Saravagi, S. Agrawal, M. Saravagi, J. M. Chatterjee, and M. Agarwal, “Diagnosis of Lumbar Spondylolisthesis Using Optimized Pretrained CNN Models,” Comput. Intell. Neurosci. , vol. 2022, p. e7459260, Apr. 2022, doi: 10.1155/2022/7459260 . F. Varçin, H. Erbay, E. Çetin, İ. Çetin, and T. Kültür, “Diagnosis of Lumbar Spondylolisthesis via Convolutional Neural Networks,” in 2019 International Artificial Intelligence and Data Processing Symposium (IDAP) , Sep. 2019, pp. 1–4. doi: 10.1109/IDAP.2019.8875988 . M. Jamaludin, “Automated analysis of spinal MRI using deep learning,” http://purl.org/dc/dcmitype/Text , University of Oxford, 2017. Accessed: Mar. 08, 2023. [Online]. Available: https://ora.ox.ac.uk/objects/uuid:c9d0e126-e44f-43d 9-abcf-140d1a73e58d A. Nagpal and G. Gabrani, “Python for Data Analytics, Scientific and Technical Applications,” in 2019 Amity International Conference on Artificial Intelligence (AICAI) , Feb. 2019, pp. 140–145. doi: 10.1109/AICAI.2019.8701341 . E. A. Kassaw, G. T. Aboye, D. Yilma, S. Dhaba, and G. L. Simegn, “The impact of khat chewing on heart activity and rehabilitation therapy from khat addiction in healthy khat chewers,” Sci. Rep. , vol. 12, no. 1, Art. no. 1, Dec. 2022, doi: 10.1038/s41598-022-26714-w . N. Ujgare and S. Baviskar, “Conversion of DICOM Image in to JPEG, BMP and PNG Image Format,” Int. J. Comput. Appl., vol. 62, pp. 22–26, Jan. 2013, doi: 10.5120/10124-4886 . N. Kumar and M. Nachamai, “Noise Removal and Filtering Techniques Used in Medical Images,” Orient. J. Comput. Sci. Technol. , vol. 10, no. 1, pp. 103–113, Mar. 2017. P. C. Nair, “Comparative Analysis of Various Denoising Techniques for MRI Images,” 2015. Accessed: Mar. 08, 2023. [Online]. Available: https://www.semanticscholar.org/paper/Comparative-Analysis-of-Various-Denoising-for-MRI-Nair/e42dd1b1f1e0d 5d65555246cba66741ff66a4498 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-2753236","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":187595049,"identity":"3a02536f-fe49-4bc6-b426-ebd32bf50a4c","order_by":0,"name":"Ewunate Assaye Kassaw","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4ElEQVRIiWNgGAWjYPCC/3b87A1A2sCCaC3MyZI9B0BaJIjXwrjhRgKIQYQW/v7jDz/zVLAxS858fnXDjwIJBv727gS8WiRu5BhL85zh4eOXzim72QN0mMSZsxvwW3ODh0Gat02CWXJ2TtoNHqAWA4lc/Frkzx9//Jv3nwHjhptn0m7+IUaLwYEEM2nehgSg99mP3SbKFsMbOWaWc44dAAZyDtttGQMJHoJ+kQM67MabmgPAqDz+7OabPzZy/O29BLwPBEw8YIrHAEwSVA4CjD/AFPsDolSPglEwCkbByAMAw0FIHgmcr3sAAAAASUVORK5CYII=","orcid":"","institution":"University of Gondar","correspondingAuthor":true,"prefix":"","firstName":"Ewunate","middleName":"Assaye","lastName":"Kassaw","suffix":""},{"id":187595052,"identity":"b27fe389-825e-4e37-99b1-0175d53168ff","order_by":1,"name":"Bekele Mulat Enyew","email":"","orcid":"","institution":"University of Gondar","correspondingAuthor":false,"prefix":"","firstName":"Bekele","middleName":"Mulat","lastName":"Enyew","suffix":""},{"id":187595055,"identity":"018fc86c-1d29-4b7b-bb83-433be0d7a693","order_by":2,"name":"Abebe Alemu Abitew","email":"","orcid":"","institution":"University of Gondar","correspondingAuthor":false,"prefix":"","firstName":"Abebe","middleName":"Alemu","lastName":"Abitew","suffix":""},{"id":187595059,"identity":"7e77bfc7-aeef-474b-973d-f5a303fb87a2","order_by":3,"name":"Yonathan Gebrewold","email":"","orcid":"","institution":"University of Gondar","correspondingAuthor":false,"prefix":"","firstName":"Yonathan","middleName":"","lastName":"Gebrewold","suffix":""}],"badges":[],"createdAt":"2023-03-29 18:14:26","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-2753236/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-2753236/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":35057917,"identity":"dc4747d2-4d00-45fb-8ec2-c958cea7b3e5","added_by":"auto","created_at":"2023-03-30 18:22:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":77981,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 2.1 architecture of proposed model\u003c/p\u003e","description":"","filename":"2.1.png","url":"https://assets-eu.researchsquare.com/files/rs-2753236/v1/ae4dad4e93f620592f31125e.png"},{"id":35058262,"identity":"76975dd1-8e4f-468a-b522-8dcab32adb3b","added_by":"auto","created_at":"2023-03-30 18:30:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":150745,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 2.2 Proposed feature extraction model\u003c/p\u003e","description":"","filename":"2.2.png","url":"https://assets-eu.researchsquare.com/files/rs-2753236/v1/4451d7a6f8615d84fb0177de.png"},{"id":35057918,"identity":"668def8f-f3a7-4e1f-8286-692612add21d","added_by":"auto","created_at":"2023-03-30 18:22:08","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":47918,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 3.1. Accuracy and loss graph for ensemble model\u003c/p\u003e","description":"","filename":"3.1.png","url":"https://assets-eu.researchsquare.com/files/rs-2753236/v1/e1357605196b7b19722e369b.png"},{"id":35058261,"identity":"e9ea38d5-9c0c-4077-8a0f-f6fbbef3a80c","added_by":"auto","created_at":"2023-03-30 18:30:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":45000,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 3.2 proposed ensemble model accuracy result for various parameters\u003c/p\u003e","description":"","filename":"3.2.png","url":"https://assets-eu.researchsquare.com/files/rs-2753236/v1/090911d5743dae00087d0059.png"},{"id":35057922,"identity":"3265307f-7758-4d64-867c-dab0fc93865e","added_by":"auto","created_at":"2023-03-30 18:22:08","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":147603,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 3.3. the proposed model test result after image preprocessing (a) accuracy loss graph (b) confusion matrix, and (c) ROC_AUC score.\u003c/p\u003e","description":"","filename":"3.3.png","url":"https://assets-eu.researchsquare.com/files/rs-2753236/v1/276d497331d449991bfd5416.png"},{"id":35057920,"identity":"6ab817ce-8120-4798-ad2c-90848779f463","added_by":"auto","created_at":"2023-03-30 18:22:08","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":269903,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 3.4. the proposed model test result after image preprocessing (a) accuracy loss graph (b) confusion matrix, and (c) ROC_AUC score.\u003c/p\u003e","description":"","filename":"3.4.png","url":"https://assets-eu.researchsquare.com/files/rs-2753236/v1/97cf4d17adcfceacd1e3431d.png"},{"id":41908535,"identity":"306f00b2-d6a3-4f6a-bfd1-ea3350012530","added_by":"auto","created_at":"2023-08-22 03:52:29","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":982965,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-2753236/v1/b535eaeb-ee78-4eba-9dbf-3ebf241de0a4.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Classification of lumbar spondylosis from MRI images using CNN ensemble method","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe number of years with a disability caused by lower back pain increased by 54% between 1990 and 2015, with the highest rise occurring in third-world nations [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Between 50% and 80% of people are thought to have lower back ache in their lives caused by intervertebral disc degeneration and damage which imposes a huge socioeconomic burden on the neighborhood [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The lumbar area is predominantly affected by spondylosis, a common disease of the spinal column, mostly due to an unfavorable ratio between the mechanical loads [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eLumbar spondylosis (LS) is becoming a significant clinical problem that the world is currently facing. Lumbar spondylosis symptoms include degeneration of the intervertebral discs or facet joints, formation of bony spurs (osteophytes), sclerosis of the vertebral body or endplates, hypertrophy of ligaments, or, in more severe cases, narrowing of the spinal line or disk space [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. A history, physical exam, and variety of imaging techniques are carried out to diagnose LS condition [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. The basis for assessing lower back pain is currently MRI rather than CT and X-ray imaging because of the availability of soft tissue features, cross-section capacity, and the absence of hazardous radiation [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Several studies have been conducted using machine learning, especially CNNs, for the LS classification and diagnostic tasks, despite the fact that the majority of them used insufficient datasets and neglected important picture preprocessing techniques [\u003cspan additionalcitationids=\"CR7 CR8 CR9 CR10 CR11\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u0026ndash;[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eColumbo software is used for segmentation but more research is needed to see whether it can genuinely improve outcomes standardization, reduce the reading time, and increase inter-rater agreement [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. According to a research team, the herniated lumbar disc can be classified as foraminal, post-lateral, or median using the VGG16 end-to-end classifier model. They did not include sagittal views, scans, or other disorders that were essential to the investigation, nor did they employ any extra pretrained models [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. An improved model was developed using the TFLite model optimization technique. The model skips crucial image processing methods like noise removal that could improve the model's performance and only uses SoftMax as a classifier [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Two separate deep neural networks, AlexNet and GoogLeNet, were utilized to diagnose two classes of spondylolisthesis. The results show that GoogleLeNet outperforms AlexNet when trained with private X-ray images. Implementing crucial strategies like feature segmentation and classification techniques wasn't demonstrated [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Previous studies that simply examined sagittal T2-weighted images were unable to accurately detect spinal canal stenosis and modic alterations, necessitating the use of both T-1 and T-2 weighted scans. While analyzing lumbar spine MRIs in clinical settings, it is also necessary to take into account automatic disc herniation or nerve root compression identification [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eHealthcare systems in underdeveloped countries have a shortage of radiologists [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Smaller data sets, the exclusion of fundamental image processing methods, and a failure to account for the use of ensemble CNN models for enhanced performance are some of the shortcomings in the machine learning models developed so far. Using an MRI dataset that had undergone considerable preprocessing, the aim of this study was to develop a deep learning model by merging CNN models to identify the various LS illness classifications.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cp\u003eAn ensemble model was developed by combining the CNN model with VGG19 to detect and diagnose Lumbar spondylosis from MRI image. To develop this ensemble model data were collected, the date was prepared, preprocessed, features extracted, and delivered to the training algorithm. The test data was then given into the classifier, whose performance was assessed using a variety of metrics, and the model was deployed using the Flask environment. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e2.1\u003c/span\u003e shows the general block diagram of the method used in this study.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Dataset preparation\u003c/h2\u003e \u003cp\u003eThe specialized hospital of the university of Gondar provided 11158 T1 and T2 weighted axial and sagittal sequence Lumbosacral MRI images that were scanned between October 27, 2021, and September 6, 2022. The data set contains 2739 normal, 2841 disc-bulge, 2804 disc-desiccation, and 2774 nerve root compression (NRC) images. Then the data were converted from DICOM to PNG format for the convenience of image processing [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The initial RGBA MRI image format is then converted to three channel format which is suitable for further processing. The required picture portions were crop using array slices, and extra annotations were minimized. To take the device's computing power into account, the 64x64 pixel resolution was then obtained. Finally, using label encoder, the data was labeled as LS0\u0026thinsp;=\u0026thinsp;bulge, LS1\u0026thinsp;=\u0026thinsp;desiccation, LS2\u0026thinsp;=\u0026thinsp;normal, and LS3\u0026thinsp;=\u0026thinsp;NRC.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Image preprocessing\u003c/h2\u003e \u003cp\u003eMRI image is susceptible to salt-and-pepper, gaussian, Rician and thermal noises which can be removed by keeping the quality and enhancing the smoothness of image using filters [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. For filters these noises gaussian, mean, median, mean-median, and adaptive histogram equalization methods are reported to performe better for MRI imges [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Following the individual application of these filters, the overall system performance was assessed as shown in Table \u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e2.1\u003c/span\u003e, and the median filter outperformed the others.\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 2.1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe performance of the model after using different filtering methods\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=\"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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eName\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eFiltering techniques\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c5\" namest=\"c3\"\u003e \u003cp\u003eResult /accuracy in%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eValidation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTesting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC ROC\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003eExperiment result after filtering preprocess\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGaussian filter\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e98.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e97.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.32\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean filter\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e98.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e97.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.13\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMedian filter\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e97.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e97.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.59\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean-median filter\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e99.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e97.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e96.72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAHE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e99.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e96.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.55\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\u003eAssumed to be the task after denoising, image segmentation was left out of this study because its implementation did not result in any improvements. To increase the training dataset and avoid overfitting rotation, shifting, adding brightness, flipping, and zooming offline data augmentation methods were used.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Model development\u003c/h2\u003e \u003cp\u003eThe dataset was successfully split into an 80:20 ratio utilizing hold out approach after the MRI images were correctly preprocessed. There were 7810, 1674, and 1674 datasets distributed for training, validation, and testing, respectively. SoftMax classifier was used to assess a few pre-trained models at epochs of 100, and the result shows VGG net performed better with an accuracy value of 93.32%. Then, features are extracted from MRI images using the suggested CNN model and proposed VGG19 net, in parallel. Then, in order to create a model with improved performance, the developed CNN classifier model and the pretrained VGG19 model was combined. The classifier receives all of the features that were retrieved from the two models in one form. The hyperparameters are modified with an Adam optimizer, categorical cross-entropy, learning rate of 0.001, at epochs of 100, and a batch size of 32. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e shows the overview of the proposed model development. The accuracy, precision, recall, F-Measure, confusion matrices, specificity, and AUC ROC curve metrics were used to assess the model's performance. Finally, a graphical user interface was developed using a flask environment to classify an input image as belonging to one of the classes and offer a probability percentage out of 100%.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. Experimental Result And Discussion","content":"\u003cdiv class=\"Section2\" id=\"Sec7\"\u003e\n \u003ch2\u003e3.1. Results\u003c/h2\u003e\n \u003cdiv class=\"Section3\" id=\"Sec8\"\u003e\n \u003ch2\u003e3.1.1. Experiment without image preprocessing\u003c/h2\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e3.1\u003c/span\u003e shows the outcomes of tests conducted using proposed CNN, pretrained VGG19, and ensemble models without using image preprocessing.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable border=\"1\" id=\"Tab2\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3.1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eExperiment result before data preprocess\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eName\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMetrics and model name\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTest accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAUC-ROC\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"3\"\u003e\n \u003cp\u003eExperiment result before filtering, segmentation and augmentation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProposed CNN model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.85%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.20%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProposed VGG19 model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e90.37%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.31%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProposed ensemble model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.44%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.10%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv class=\"Section3\" id=\"Sec9\"\u003e\n \u003ch2\u003e3.1.2. Experiment result for various parameters\u003c/h2\u003e\n \u003cp\u003eA comprehensive experiment has been carried out a to determine the best parameters. Figure 3 .2 shows an overview of comparisons for a few parameters.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"Section3\" id=\"Sec10\"\u003e\n \u003ch2\u003e3.1.3. Final model detail experimental result discussion\u003c/h2\u003e\n \u003cp\u003eThe proposed ensemble model was found to be the most ideal model at 100 epochs for classification of lumbar spondylosis following augmentation as shown in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3.2\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable border=\"1\" id=\"Tab3\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3.2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eExperimental result for final model\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eName\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eModel name\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eResult %\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTraining\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003evalidation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003etesting\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eExperiment result for final ensemble model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProposed ensemble\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e98.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e98.16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e3.3\u003c/span\u003e shows the results of testing the proposed ensemble method using the preprocessed image. The precision, recall and F1-score results are also shown in Table \u003cspan class=\"InternalRef\"\u003e3.3\u003c/span\u003e.\u0026nbsp;\u003c/p\u003e\n \u003ctable border=\"1\" id=\"Tab4\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3.3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003ePrecision, Recall and F1-score results of the proposed ensemble method\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1-Score\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBulge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDesiccation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNormal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNRC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"Section3\" id=\"Sec11\"\u003e\n \u003ch2\u003e3.1.4. Model Deployment\u003c/h2\u003e\n \u003cp\u003eThe deployment overview of the model user interface is shown in Fig. 3.4.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv class=\"Section2\" id=\"Sec12\"\u003e\n \u003ch2\u003e3.2. Discussion\u003c/h2\u003e\n \u003cp\u003eIn this study ensemble method developed by concatenating proposed CNN and VGG19 was used to classify lumbar spondylosis. Initially the performance of the proposed CNN, VGG19 and ensemble models were tested and scored an accuracy result of 91.85%, 90.37%, and 94.44% respectively. The accuracy vs. epoch and loss vs. epoch plots in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3.1\u003c/span\u003e shows that accuracy increases and loss decreases as epoch increases and epoch size 100 is an ideal for the proposed model. Moreover, the proposed ensemble model accuracy was tested and resulted better result for train test split of 80/20, batch size of 32, learning rate of 0.001, dropout of 0.1, and max-pooling mechanism as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3.1\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eTo improve the performance further image preprocessing and augmentation processes were done on the input MRI data. When tested using the confusion matrix provided in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3.3\u003c/span\u003e, the proposed model achieved a weighted average value of 98.16%, 98%, 98%, and 98% for accuracy, precision recall, and F1-score, respectively. This outcome demonstrates how preprocessing and augmentation techniques significantly increased the model\u0026apos;s accuracy. The model accuracy loss graphs and the AUC ROC curve also demonstrate how preprocessing and augmentation considerably enhanced the performance of the proposed model. On test data, the AUC ROC score was 98.78%, indicating the model can classify between class labels accurately. A convenient environment for real-time use was created by the deployment of the built model.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eOne of the issues the globe is facing is lumbar spondylosis (LS), which is becoming into a severe clinical concern. Due to the availability of soft tissue information, cross-sectional capacity, and lack of hazardous radiation, MRI is now the primary method used to study LS. The row MRI image was augmented and preprocessed, and the proposed CNN and VGG19 models concatenated to develop an ensemble model. Testing the ensemble model's accuracy resulted in a 98.16% accuracy score, which is a promising result for using it to diagnose lumbar spondylosis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003col\u003e\n\u003cli\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eEthical approval letters were obtained from Institutional review boards of University of Gondar institute of health science. All methods were carried out in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards.\u003c/p\u003e\n\u003col start=\"2\"\u003e\n\u003cli\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThis research did not involve humans, animals, or other subjects\u003c/p\u003e\n\u003col start=\"3\"\u003e\n\u003cli\u003e\u003cstrong\u003eAvailability of data and material\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study are available from the corresponding author on reasonable request\u003c/p\u003e\n\u003col start=\"4\"\u003e\n\u003cli\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003col start=\"5\"\u003e\n\u003cli\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003col start=\"6\"\u003e\n\u003cli\u003e\u003cstrong\u003eAuthors' contributions\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eBM and EA conceptualized, designed, and implemented in collaboration with the co-investigators AA and YG. All authors contributed to the preliminary study, the design, prototyping, and testing. The article was drafted by BM and EA, taking into account the comments and suggestions of the co-authors. All co-authors had the opportunity to comment on the manuscript and approved the final version for publication.\u003c/p\u003e\n\u003col start=\"7\"\u003e\n\u003cli\u003e\u003cstrong\u003eAcknowledgments \u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe specialized hospital at the University of Gondar provided the MRI data sets needed to complete the investigation. Much assistance was provided by the Department of Information Technology at University of Gondar.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003e\u003cspan\u003eM. Kahere and T. Ginindza, \u0026ldquo;Mapping evidence on the prevalence, incidence, risk factors and cost associated with chronic low back pain among adults in Sub-Saharan Africa: a systematic scoping review protocol,\u0026rdquo; Syst. Rev., vol. 9, no. 1, p. 57, Mar. 2020, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13643-020-01321-w\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eM. Kolenkiewicz, A. Włodarczyk, and J. Wojtkiewicz, \u0026ldquo;Diagnosis and Incidence of Spondylosis and Cervical Disc Disorders in the University Clinical Hospital in Olsztyn, in Years 2011\u0026ndash;2015,\u0026rdquo; \u003cem\u003eBioMed Res. Int.\u003c/em\u003e, vol. 2018, p.\u0026nbsp;5643839, Mar. 2018, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1155/2018/5643839\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eR. Tsujimoto \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Prevalence of lumbar spondylosis and its association with low back pain among community-dwelling Japanese women,\u0026rdquo; BMC Musculoskelet. Disord., vol.\u0026nbsp;17, no. 1, p.\u0026nbsp;493, Dec. 2016, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12891-016-1343-x\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eC. Barlotta, \u0026ldquo;Prevalence and Risk Factors for Lumbar Spondylosis and Its Association with Low Back Pain among Rural Korean Residents.,\u0026rdquo; \u003cem\u003eMinerva Stomatol.\u003c/em\u003e, vol.\u0026nbsp;15, no. 6, pp.\u0026nbsp;458\u0026ndash;461, Jun. 1966.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eM. Munir Mir, A. John, M. Naeem, H. Butt, and A. Ali, \u0026ldquo;Prevalence and Radiological Evaluation of Lumbar Spondylosis on Magnetic Resonance Imaging,\u0026rdquo; \u003cem\u003eEAS J. Radiol. Imaging Technol.\u003c/em\u003e, vol.\u0026nbsp;3, pp.\u0026nbsp;57\u0026ndash;65, Apr. 2021, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.36349/easjrit.2021.v03i02.005\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eN. C. Lehnen \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Detection of Degenerative Changes on MR Images of the Lumbar Spine with a Convolutional Neural Network: A Feasibility Study,\u0026rdquo; \u003cem\u003eDiagnostics\u003c/em\u003e, vol.\u0026nbsp;11, no. 5, Art. no. 5, May 2021, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/diagnostics11050902\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eA. Hirschmann, J. Cyriac, B. Stieltjes, T. Kober, J. Richiardi, and P. Omoumi, \u0026ldquo;Artificial Intelligence in Musculoskeletal Imaging: Review of Current Literature, Challenges, and Trends,\u0026rdquo; \u003cem\u003eSemin. Musculoskelet. Radiol.\u003c/em\u003e, vol.\u0026nbsp;23, no. 03, pp.\u0026nbsp;304\u0026ndash;311, Jun. 2019, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1055/s-0039-1684024\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eW. Mbarki, M. Bouchouicha, S. Frizzi, F. Tshibasu, L. B. Farhat, and M. Sayadi, \u0026ldquo;Lumbar spine discs classification based on deep convolutional neural networks using axial view MRI,\u0026rdquo; Interdiscip. Neurosurg., vol.\u0026nbsp;22, p.\u0026nbsp;100837, Dec. 2020, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.inat.2020.100837\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eD. Saravagi, S. Agrawal, M. Saravagi, J. M. Chatterjee, and M. Agarwal, \u0026ldquo;Diagnosis of Lumbar Spondylolisthesis Using Optimized Pretrained CNN Models,\u0026rdquo; \u003cem\u003eComput. Intell. Neurosci.\u003c/em\u003e, vol. 2022, p. e7459260, Apr. 2022, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1155/2022/7459260\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eF. Var\u0026ccedil;in, H. Erbay, E. \u0026Ccedil;etin, İ. \u0026Ccedil;etin, and T. K\u0026uuml;lt\u0026uuml;r, \u0026ldquo;Diagnosis of Lumbar Spondylolisthesis via Convolutional Neural Networks,\u0026rdquo; in \u003cem\u003e2019 International Artificial Intelligence and Data Processing Symposium (IDAP)\u003c/em\u003e, Sep. 2019, pp.\u0026nbsp;1\u0026ndash;4. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/IDAP.2019.8875988\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eM. Jamaludin, \u0026ldquo;Automated analysis of spinal MRI using deep learning,\u0026rdquo; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://purl.org/dc/dcmitype/Text\u003c/span\u003e\u003c/span\u003e, University of Oxford, 2017. Accessed: Mar. 08, 2023. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ora.ox.ac.uk/objects/uuid:c9d0e126-e44f-43d\u003c/span\u003e\u003c/span\u003e9-abcf-140d1a73e58d\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eA. Nagpal and G. Gabrani, \u0026ldquo;Python for Data Analytics, Scientific and Technical Applications,\u0026rdquo; in 2019 \u003cem\u003eAmity International Conference on Artificial Intelligence (AICAI)\u003c/em\u003e, Feb. 2019, pp.\u0026nbsp;140\u0026ndash;145. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/AICAI.2019.8701341\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eE. A. Kassaw, G. T. Aboye, D. Yilma, S. Dhaba, and G. L. Simegn, \u0026ldquo;The impact of khat chewing on heart activity and rehabilitation therapy from khat addiction in healthy khat chewers,\u0026rdquo; \u003cem\u003eSci. Rep.\u003c/em\u003e, vol.\u0026nbsp;12, no. 1, Art. no. 1, Dec. 2022, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-022-26714-w\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eN. Ujgare and S. Baviskar, \u0026ldquo;Conversion of DICOM Image in to JPEG, BMP and PNG Image Format,\u0026rdquo; Int. J. Comput. Appl., vol.\u0026nbsp;62, pp.\u0026nbsp;22\u0026ndash;26, Jan. 2013, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5120/10124-4886\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eN. Kumar and M. Nachamai, \u0026ldquo;Noise Removal and Filtering Techniques Used in Medical Images,\u0026rdquo; \u003cem\u003eOrient. J. Comput. Sci. Technol.\u003c/em\u003e, vol.\u0026nbsp;10, no. 1, pp.\u0026nbsp;103\u0026ndash;113, Mar. 2017.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eP. C. Nair, \u0026ldquo;Comparative Analysis of Various Denoising Techniques for MRI Images,\u0026rdquo; 2015. Accessed: Mar. 08, 2023. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.semanticscholar.org/paper/Comparative-Analysis-of-Various-Denoising-for-MRI-Nair/e42dd1b1f1e0d\u003c/span\u003e\u003c/span\u003e5d65555246cba66741ff66a4498\u003c/span\u003e\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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":"CNN, ensemble technique, image processing, lumbar spondylosis, MRI","lastPublishedDoi":"10.21203/rs.3.rs-2753236/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-2753236/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Due to an unfavorable ratio between the mechanical load and the size of the intervertebral discs, lumbar spondylosis, one of the most common causes of morbidity and disability. The preferred imaging technique for determining the origins of complex lower back pain is MRI. Healthcare systems in underdeveloped countries have a shortage of radiologists. Developing a CNN ensemble model for diagnosing lumbar spondylosis from MRI images was the aim of this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e: 11158 T1 and T2 labeled MRI scans were collected from the University of Gondar specialized hospital and prepared for image processing. Since the median filter performed better than the others, it was chosen to denoise the data. The data was then augmented and split into an 80:20 train test ratio. An ensemble model was built by concatenating the proposed CNN and VGG19 models. Finally, the model was deployed.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: An ensemble model achieved strong performance of 98.16% accuracy, 98% recall, and 98% precision. The GUI provides a setting appropriate for routine model usage.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e: The research confirms that lumbar spondylosis can be diagnosed using MRI data and a CNN ensemble model.\u003c/p\u003e","manuscriptTitle":"Classification of lumbar spondylosis from MRI images using CNN ensemble method","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2023-03-30 18:22:03","doi":"10.21203/rs.3.rs-2753236/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":"3a7b6e4f-e427-4668-91fa-ae238d1bdd6f","owner":[],"postedDate":"March 30th, 2023","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":20310026,"name":"Physical sciences/Engineering/Biomedical engineering"},{"id":20310027,"name":"Physical sciences/Engineering/Electrical and electronic engineering"},{"id":20310028,"name":"Physical sciences/Engineering"}],"tags":[],"updatedAt":"2023-08-22T03:44:22+00:00","versionOfRecord":[],"versionCreatedAt":"2023-03-30 18:22:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-2753236","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-2753236","identity":"rs-2753236","version":["v1"]},"buildId":"J0_U0BvcaRcwD8yVFaRlm","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}