Deep Learning-based Automated Identification of Knee Arthroplasty Implants

Deep Learning-based Automated Identification of Knee Arthroplasty Implants | 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 Deep Learning-based Automated Identification of Knee Arthroplasty Implants Zibo Gong, Xiaoxin He, Ming He, Xinzhe Fu, Yonghui Fu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4299072/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: A critical step in preoperative planning for revision knee arthroplasty is the identification of the failed implant. The purpose of this study was to develop and evaluate the performance of deep learning methods based on convolutional neural networks (CNN) to detect and identify specific knee arthroplasty models. Methods: In this study, we propose a novel deep learning-based approach to identify knee arthroplasty implants’ design using both anterior-posterior (AP) and lateral images. We harness the pre-trained ResNet50 CNN model and employ transfer learning methods to adapt the model for implants identification task using a total of 1814 radiographs of 8 different knee arthroplasty implant designs. Performance was compared with operative note and crosschecked with implant sheets. We also evaluate the difference of performance of models trained with AP images, lateral images or both. Results: The training and validation data sets were comprised of 907 AP and 907 lateral view radiographs across 733 patients and included 8 knee arthroplasty implants from 5 leading implant manufactures. After 1000 training epochs the model classified 8 implant models with very high accuracy. Our results showed that jointly using AP images and lateral images improved the classification accuracy of the CNN model. Conclusion: CNN can accurately distinguish between specific knee arthroplasty designs. This technology could offer a useful adjunct to the surgeon in preoperative identification of the prior implant. Furthermore, using both AP images and lateral images to train the CNN is more effective than using images from only one perspective. Health sciences/Diseases Health sciences/Medical research Machine learning Deep learning Artificial intelligence Knee arthroplasty Neural networks Figures Figure 1 Figure 2 Figure 3 Introduction Osteoarthritis (OA) is a major cause of disability worldwide, accounting for 40–60% of patients with degenerative diseases of the musculoskeletal system, and thus carries high socioeconomic costs. Total knee arthroplasty (TKA) is a successful treatment of choice for patients with end-stage OA. However, despite the clinical effectiveness of TKAs, the number of revision TKAs has increased with time. Projections based on population studies point to continued increases in the prevalence of revision surgeries[ 1 ]. One of the critical steps in preoperative planning for revision TKA is to identify the failed prosthesis. A recent survey of arthroplasty surgeons demonstrated that it took roughly 20 minutes for surgeons to identify the implant preoperatively for each revision procedure. Failure rate of implants identification could be 10% preoperatively, with 2% of implants not being identified intraoperatively. Failure to identify implant preoperatively resulted in additional requested implants, added surgical time, increased perioperative morbidity, increased healthcare cost[ 2 ]. Machine learning is an application of artificial intelligence (AI). It enables computers to find hidden insights without being explicitly programmed using algorithms that iteratively learn from the data. In the field of general imaging and computer vision, deep learning is the leading machine learning tool. Deep learning refers to techniques that build on developments in artificial neural networks in which multiple network layers are added to increase the levels of abstraction and performanc[ 3 ]. Recently, deep learning, and specifically convolution neural networks (CNN), has demonstrated groundbreaking outcomes in plenty of general image recognition and computer-aided diagnosis tasks[ 4 ]. Within orthopedics, these powerful models already can reach human level performance in diagnosis of fractures[ 5 ] and staging knee OA severity[ 6 ], which clearly forecasts their possible utilization in clinical practice in the near future. We hypothesized that deep learning-based AI algorithms could facilitate the automated identification of knee arthroplasty implants, thereby aiding in preoperative planning and resulting in saved time and healthcare cost spent on this labor-intense task. Methods Study Design and Radiograph Acquisition This study was approved by the ethics committee of Shengjing Hospital of China Medical University (2021PS550K), and all the experiments were performed in accordance with relevant guidelines and regulations. Informed consent from patients was waived because of the retrospective nature of the study and the minimal risk involved. All the X-ray images taken between June 1, 2011 and Dec 1, 2020 at one university hospital were collected retrospectively. We collected all radiographs from operations performed by 3 senior arthroplasty surgeons to make sure a wide range of prosthesis manufacturers and prosthesis designs to be included. During the time of image collection, all identifying information was removed. We checked the primary operative records and crosschecked with implant sheets to confirm the implant type. We included implant designs only if more than 10 images per model were identified[6]. In this study we included a total of 1814 images from the medical records of 733 patients. Zimmer and Biomet were treated as two different manufacturers. The following 8 designs from the 5 manufacturers were included: Depuy Attune (Depuy Synthes, Warsaw, IN), Depuy PFC Sigma RP (Depuy Synthes, Warsaw, IN), LINK GEMINI MK II, LINK GEMINI PS, Biomet Vanguard (Zimmer Biomet, Warsaw, IN), Zimmer Nexgen (Zimmer Biomet, Warsaw, IN), Smith and Nephew Genesis II (Smith & Nephew Incorporated, Andover, MA), and Biomet Oxford (Zimmer Biomet, Warsaw, IN). We excluded implant designs that were less than 10 images per model. Figure 1 demonstrated example of anterior-posterior (AP) and lateral images of each implant design[6]. Overview of Framework We used convolutional neural network-based algorithms (CNN) for classification of knee implants. As a CNN model typically involved millions of parameters while our data set only contained about one thousand images, it was infeasible to train a CNN model from scratch using our data. Therefore, we adopted the transfer learning framework and the ResNet architecture[7]. It started with a ResNet50 network pre-trained on ImageNet[8], and did gradient descent (backward-propagation) using our training data only on the last two layers of the network. We trained one network using the AP images, and another network using the lateral images. Finally, we used a two-layer feed-forward network to combine the output of AP-trained network and lateral-trained network to evaluate whether jointly considering AP images and lateral images could improve the classification accuracy. Figure 2 demonstrates the overview of the framework of our deep learning-based method. Dataset: Our dataset contained 1814 images from 8 different kind of implants Image Preprocessing: We followed standard procedures to pre-process our training data so that it could work with a network trained on ImageNet. We rescaled each image to a size of 224*224, and normalized it according to ImageNet standards. We also performed data augmentation, i.e., random rotation, horizontal flips, etc., to increase the amount of training data and make our algorithm robust to the orientation of the images. Dataset Division We first divided the set of patients into three groups of sizes ~ 60% (group 1), ~ 20% (group 2), and ~ 20% (group 3). This split technique was used on a per-design basis to ensure the ratio of each implant remained constant. Next, we used the AP images and lateral images of patients in group 1 for training, those of patients in group 2 for validation, and those of patients in group 3 for testing. The validation set was used to compute cross-validation loss for hyper-parameter tuning and early stopping determination. Model Training We adopted the adaptive gradient method ADAM[9] to train our models. Based on the cross-validation loss, we chose the hyper-parameters for ADAM as (learning rate \({\alpha }\) = 0.001, \({{\beta }}_{1}=0.9, {\beta }_{2}=0.99, ϵ={10}^{-8},\) weight_decay = 0). The maximum number of epochs was 1000 and the batch size was 16. The early stopping threshold was set to 8. During the training process of each network, the early stopping threshold was hit after around 50 epochs. As we mentioned above, we trained one network with the AP images and another with the lateral images. We further combined the two models to see if jointly utilizing AP and lateral images could yield superior testing results. The combination was done via the following logistic-regression based method. We collected the outputs of the AP network and the lateral network (both are of the form of an 8-dimensional vector, with each element corresponding to the classification weight the network gives to the category of implants), and then fed them as the input to a two-layer feed-forward neural network, and trained the network with the data from the validation set. Note that the above construction relied on our dataset division procedure, where the training set, validation set, and testing set, each contained the AP and lateral images of the same set of patients. We referred to the resulting network constructed from the outputs of AP networks and Lateral network as the “joint network”. Model Testing We tested our models (AP, Lateral, Joint) using the testing set. The prediction result on each testing image was an 8-dimensional vector, with each coordinate representing the classification confidence of the corresponding category of implants. Statistical Analysis Since we were studying a multi-class classification problem, we would directly present the confusion matrices of our methods on the testing data and compute the operation characteristics generalized for multi-class classification. Results Training Progress The following figure (Fig. 3) showed the training progress of our method (we used the AP-network as an example). As the training proceeded, the network adjusted its parameters and learned the correct classification function as demonstrated by decrease in both training and validation loss. Testing Performance In this section, we evaluated our neural network models on the testing data sets. In the testing sets, we had in total 177 images. The classification results of the three neural network models (AP, Lateral, Joint) were presented in the form of tables (Table 1–3). The column labels of each table indicated the ground-truth implant type, while row labels indicated the classification results. For example, the first row in the table of the AP-network indicated that the AP network classified 2 images with type-A plant as type A, and 1 image with type-A plant as type B. AP-Network The classification results (Confusion Matrix) of the AP-Network were shown in Table 1. The AP-network achieved an overall classification accuracy of 92.7%. Although the overall accuracy was high, the accuracy on some implants with scare training data was less satisfying. For example, its classification accuracy on implant type-G was about 58.3%. Table 1: Classification Results (Confusion Matrix) of the AP-Network Lateral-Network The classification results (Confusion Matrix) of the Lateral-Network were shown in Table 2. The Lateral network also achieved a high accuracy of 91.5%. It outperformed the AP-network on certain implant types (e.g., type G, with an accuracy of ~ 91.7%) but underperformed on some other types (e.g., type E). Table 2: Classification Results (Confusion Matrix) of the Lateral-Network Joint Network The classification results (Confusion Matrix) of the Joint Network were shown in Table 3. The join-network that essentially combined the AP-network and the Lateral-network had an overall accuracy of 95.4%. In addition to its higher overall accuracy, it also achieved great classification accuracy on each individual implant type. Essentially, from the experiments, we saw that combining AP images with lateral images had the potential to achieve the best of both worlds, and improved the classification performance of only using one category of image data. Table 3: Classification Results (Confusion Matrix) of the Joint Network Operation Characteristics For a multi-class classification method, let \(M\) be its confusion matrix on the testing data. We used \({M}_{ij}\) to denote the entry on the \(i\)th row and \(j\)th column of \(M\). Then, for each row (class), the recall and precision of the algorithm were defined as: $$Precisio{n}_{i}=\frac{{M}_{ii}}{{\sum }_{j}{M}_{ji}}$$ $$Recal{l}_{i}=\frac{{M}_{ii}}{{\sum }_{j}{M}_{ij}}$$ The precision (recall) of the algorithm was the average of its precision (recall) for each class. Following the above definition, we had the following operating characteristics of the AP-network, Lateral-network and Joint-network (Table 4). Table 4 Operation Characteristics of Our Methods Precision Recall AP-network 0.983 0.802 Lateral-network 0.961 0.849 Joint-network 0.987 0.887 Discussion Preoperative identification of arthroplasty implants prior to revision surgery is a difficult task and an essential step in preventing increases in perioperative morbidity and increased healthcare costs. The key finding of this study was that first, convolutional neural networks could be trained to provide an automated identification of knee arthroplasty implant from radiographic images with very high accuracy. Second, training the neural networks with both AP and lateral images further improve the accuracy compared with training with one kind of images. To our knowledge, there are only 2 previous studies that have used deep learning for identifying knee arthroplasty implants from plain radiographs. One study by Yi et al.[ 10 ] used a small sample sized dataset (273 AP knee radiographs) to distinguish between TKA and UKA, and to distinguish between only two TKA models. Karnuta et al. [ 11 ]reported in the other study that their deep learning model using 682 AP knee radiographs could differentiate successfully between 9 unique knee arthroplasty implants. However, they only evaluated AP knee radiographs, as opposed to clinical practice, where lateral radiographs should always be needed, especially in differentiating cruciate retaining vs. PCL-sacrificing implants. In our study, the deep learning model developed from over 1000 AP and lateral view of knee radiographs represents the first advanced artificial intelligence model in classifying knee arthroplasty implants from radiographs. This study has several limitations. First, the data come from only one institution, the number of images for each brand implant was highly varied because the contractual relationship between the single institution and different manufactures has been changing over the past 10 years. This could result in an imbalance in trained implant data. Second, our algorithm was trained with only 8 knee arthroplasty implants, therefore, the data was limited to these models alone and should not be generalized to identify other models. Third, deep-learning based methods require large-scale training data sets. Collecting and labeling such data sets requires the combined effort of a team involving a surgeon, radiologist, and computer scientist. Also, the trained CNN functions more like a black box and is not as humanly interpretable, which might be a drawback in clinical practice. In conclusion, the technology behind machine learning is exponentially advancing and the enthusiasm for AI in healthcare is growing. With this study, we have demonstrated that a deep learning algorithm can identify the design of 8 different knee arthroplasty implants from AP and lateral knee radiograph. These findings suggest that this technology has the potential to classify knee implants prior to revision surgery, thus saving significant time, and reducing perioperative morbidity and healthcare cost. It is hoped that it can be used to collect large-scale implant information and may be applied to mobile phone applications in the future. Declarations Author Contribution Z.G. and X.H. contributed equally to this work. Z.G.: Resources, Investigation, Data curation, Writing—original draf. X.H.: Data curation, Project administration, Writing—review & editing M.H.: Data curation, Resources, Writing—review & editing X.F.: Software, Formal analysis, Writing—review & editing.Y.F.: Conceptualization, Methodology, Writing—review & editing, Supervision. Data Availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. References Bozic KJ, Kamath AF, Ong K et al: Comparative Epidemiology of Revision Arthroplasty: Failed THA Poses Greater Clinical and Economic Burdens Than Failed TKA. Clin Orthop Relat Res 2015, 473(6):2131-2138.10.1007/s11999-014-4078-8 Wilson NA, Jehn M, York S, Davis CM, 3rd: Revision total hip and knee arthroplasty implant identification: implications for use of Unique Device Identification 2012 AAHKS member survey results. J Arthroplasty 2014, 29(2):251-255.10.1016/j.arth.2013.06.027 LeCun Y, Bengio Y, Hinton G: Deep learning. Nature 2015, 521(7553):436-444.10.1038/nature14539 Esteva A, Kuprel B, Novoa RA et al: Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017, 542(7639):115-118.10.1038/nature21056 Kim DH, MacKinnon T: Artificial intelligence in fracture detection: transfer learning from deep convolutional neural networks. Clin Radiol 2018, 73(5):439-445.10.1016/j.crad.2017.11.015 Abedin J, Antony J, McGuinness K et al: Predicting knee osteoarthritis severity: comparative modeling based on patient's data and plain X-ray images. Sci Rep 2019, 9(1):5761.10.1038/s41598-019-42215-9 He K, Zhang X, Ren S, Sun J: Identity mappings in deep residual networks. In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14: 2016: Springer; 2016: 630–645 Krizhevsky A, Sutskever I, Hinton GE: ImageNet classification with deep convolutional neural networks. Communications of the ACM 2017, 60(6):84-90.10.1145/3065386 Kingma DP, Ba J: Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980 2014 Yi PH, Wei J, Kim TK et al: Automated detection & classification of knee arthroplasty using deep learning. Knee 2020, 27(2):535-542.10.1016/j.knee.2019.11.020 Karnuta JM, Luu BC, Roth AL et al: Artificial Intelligence to Identify Arthroplasty Implants From Radiographs of the Knee. J Arthroplasty 2021, 36(3):935-940.10.1016/j.arth.2020.10.021 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-4299072","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":300386106,"identity":"2fe90650-0b1a-4afa-91da-9cb367e22b56","order_by":0,"name":"Zibo Gong","email":"","orcid":"","institution":"Shengjing hospital of China Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zibo","middleName":"","lastName":"Gong","suffix":""},{"id":300386109,"identity":"46ee186d-4a3a-4dd6-b271-2e10bfa99855","order_by":1,"name":"Xiaoxin He","email":"","orcid":"","institution":"The Affiliated Center Hospital of Shenyang Medical College","correspondingAuthor":false,"prefix":"","firstName":"Xiaoxin","middleName":"","lastName":"He","suffix":""},{"id":300386112,"identity":"93c88c51-6b42-42d1-993f-c09f40d9265e","order_by":2,"name":"Ming He","email":"","orcid":"","institution":"Shengjing hospital of China Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ming","middleName":"","lastName":"He","suffix":""},{"id":300386115,"identity":"9321e30f-68e0-4ca4-89bd-8c8d605b06bc","order_by":3,"name":"Xinzhe Fu","email":"","orcid":"","institution":"Lab for Information and Decision Systems, Massachusetts Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"Xinzhe","middleName":"","lastName":"Fu","suffix":""},{"id":300386118,"identity":"e303efd2-5f9b-4ca6-a4bd-19688a7cd7ce","order_by":4,"name":"Yonghui Fu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA30lEQVRIie3PsQrCMBCA4YNAusR2vaKgjxAJFEHQhyk4BcFFEEELQqf6AL5IZ4tDl9i54qJ2FEHxBWy7ubQdBfNvF+6DC4BO96MRALTACMqBmU2J7TEFsM8JbUjA9lCWBGoJj7eH+8wfID89ouwlRx0K5HpLq4hKJsOdj8jPU5dHoZsfRoWQVSSVjmj5uOJn6WAUkpww2m5C8sNUQdaNiMhKkrKCHOqJrZRDWIJoKyn4MYwZJTV/MeNAvNl8hWas+pdFuBxbxuaaVZHeHih+P5GK9aKuB+RZs6PT6XT/3gcixEVxJQR8cAAAAABJRU5ErkJggg==","orcid":"","institution":"Shengjing hospital of China Medical University","correspondingAuthor":true,"prefix":"","firstName":"Yonghui","middleName":"","lastName":"Fu","suffix":""}],"badges":[],"createdAt":"2024-04-21 01:56:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4299072/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4299072/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":56283190,"identity":"381203e0-dcdb-4ca8-a405-366f731097c6","added_by":"auto","created_at":"2024-05-10 21:41:23","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":83653,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e1\u003c/strong\u003e demonstrated example of anterior-posterior (AP) and lateral radiographs of each included implant design.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4299072/v1/4ede9aa3b0b24ff1db000e8f.jpg"},{"id":56283111,"identity":"9629dd38-3b37-4101-b006-dc15fe2301fd","added_by":"auto","created_at":"2024-05-10 21:40:57","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":66194,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the framework of our deep learning-based method\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4299072/v1/d92c18f9dfb9043ec847db1c.jpg"},{"id":56283174,"identity":"ff50e02c-c9ee-4991-b4a3-cec95340b50d","added_by":"auto","created_at":"2024-05-10 21:41:17","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":27512,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and Validation Losses Curve of the AP-network\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4299072/v1/6e44b4642bbb433262afbbb2.jpg"},{"id":58557600,"identity":"5ec76c40-4b30-46a4-be7d-b1300a96b662","added_by":"auto","created_at":"2024-06-18 08:24:01","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":563077,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4299072/v1/5c84e091-66c1-4310-8b1f-c28c43e94985.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning-based Automated Identification of Knee Arthroplasty Implants","fulltext":[{"header":"Introduction","content":"\u003cp\u003eOsteoarthritis (OA) is a major cause of disability worldwide, accounting for 40\u0026ndash;60% of patients with degenerative diseases of the musculoskeletal system, and thus carries high socioeconomic costs. Total knee arthroplasty (TKA) is a successful treatment of choice for patients with end-stage OA. However, despite the clinical effectiveness of TKAs, the number of revision TKAs has increased with time. Projections based on population studies point to continued increases in the prevalence of revision surgeries[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. One of the critical steps in preoperative planning for revision TKA is to identify the failed prosthesis. A recent survey of arthroplasty surgeons demonstrated that it took roughly 20 minutes for surgeons to identify the implant preoperatively for each revision procedure. Failure rate of implants identification could be 10% preoperatively, with 2% of implants not being identified intraoperatively. Failure to identify implant preoperatively resulted in additional requested implants, added surgical time, increased perioperative morbidity, increased healthcare cost[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMachine learning is an application of artificial intelligence (AI). It enables computers to find hidden insights without being explicitly programmed using algorithms that iteratively learn from the data. In the field of general imaging and computer vision, deep learning is the leading machine learning tool. Deep learning refers to techniques that build on developments in artificial neural networks in which multiple network layers are added to increase the levels of abstraction and performanc[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Recently, deep learning, and specifically convolution neural networks (CNN), has demonstrated groundbreaking outcomes in plenty of general image recognition and computer-aided diagnosis tasks[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Within orthopedics, these powerful models already can reach human level performance in diagnosis of fractures[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] and staging knee OA severity[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], which clearly forecasts their possible utilization in clinical practice in the near future.\u003c/p\u003e \u003cp\u003eWe hypothesized that deep learning-based AI algorithms could facilitate the automated identification of knee arthroplasty implants, thereby aiding in preoperative planning and resulting in saved time and healthcare cost spent on this labor-intense task.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\"\u003e\n \u003ch2\u003eStudy Design and Radiograph Acquisition\u003c/h2\u003e\n \u003cp\u003eThis study was approved by the ethics committee of Shengjing Hospital of China Medical University (2021PS550K), and all the experiments were performed in accordance with relevant guidelines and regulations. Informed consent from patients was waived because of the retrospective nature of the study and the minimal risk involved. All the X-ray images taken between June 1, 2011 and Dec 1, 2020 at one university hospital were collected retrospectively. We collected all radiographs from operations performed by 3 senior arthroplasty surgeons to make sure a wide range of prosthesis manufacturers and prosthesis designs to be included. During the time of image collection, all identifying information was removed. We checked the primary operative records and crosschecked with implant sheets to confirm the implant type. We included implant designs only if more than 10 images per model were identified[6]. In this study we included a total of 1814 images from the medical records of 733 patients. Zimmer and Biomet were treated as two different manufacturers. The following 8 designs from the 5 manufacturers were included: Depuy Attune (Depuy Synthes, Warsaw, IN), Depuy PFC Sigma RP (Depuy Synthes, Warsaw, IN), LINK GEMINI MK II, LINK GEMINI PS, Biomet Vanguard (Zimmer Biomet, Warsaw, IN), Zimmer Nexgen (Zimmer Biomet, Warsaw, IN), Smith and Nephew Genesis II (Smith \u0026amp; Nephew Incorporated, Andover, MA), and Biomet Oxford (Zimmer Biomet, Warsaw, IN). We excluded implant designs that were less than 10 images per model. Figure\u0026nbsp;1 demonstrated example of anterior-posterior (AP) and lateral images of each implant design[6].\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\"\u003e\n \u003ch2\u003eOverview of Framework\u003c/h2\u003e\n \u003cp\u003eWe used convolutional neural network-based algorithms (CNN) for classification of knee implants. As a CNN model typically involved millions of parameters while our data set only contained about one thousand images, it was infeasible to train a CNN model from scratch using our data. Therefore, we adopted the transfer learning framework and the ResNet architecture[7]. It started with a ResNet50 network pre-trained on ImageNet[8], and did gradient descent (backward-propagation) using our training data only on the last two layers of the network. We trained one network using the AP images, and another network using the lateral images. Finally, we used a two-layer feed-forward network to combine the output of AP-trained network and lateral-trained network to evaluate whether jointly considering AP images and lateral images could improve the classification accuracy. Figure 2 demonstrates the overview of the framework of our deep learning-based method.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\"\u003e\n \u003ch2\u003eDataset:\u003c/h2\u003e\n \u003cp\u003eOur dataset contained 1814 images from 8 different kind of implants\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\"\u003e\n \u003ch2\u003eImage Preprocessing:\u003c/h2\u003e\n \u003cp\u003eWe followed standard procedures to pre-process our training data so that it could work with a network trained on ImageNet. We rescaled each image to a size of 224*224, and normalized it according to ImageNet standards. We also performed data augmentation, i.e., random rotation, horizontal flips, etc., to increase the amount of training data and make our algorithm robust to the orientation of the images.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\"\u003e\n \u003ch2\u003eDataset Division\u003c/h2\u003e\n \u003cp\u003eWe first divided the set of patients into three groups of sizes\u0026thinsp;~\u0026thinsp;60% (group 1), ~\u0026thinsp;20% (group 2), and ~\u0026thinsp;20% (group 3). This split technique was used on a per-design basis to ensure the ratio of each implant remained constant. Next, we used the AP images and lateral images of patients in group 1 for training, those of patients in group 2 for validation, and those of patients in group 3 for testing. The validation set was used to compute cross-validation loss for hyper-parameter tuning and early stopping determination.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\"\u003e\n \u003ch2\u003eModel Training\u003c/h2\u003e\n \u003cp\u003eWe adopted the adaptive gradient method ADAM[9] to train our models. Based on the cross-validation loss, we chose the hyper-parameters for ADAM as (learning rate \\({\\alpha }\\) = 0.001, \\({{\\beta }}_{1}=0.9, {\\beta }_{2}=0.99, ϵ={10}^{-8},\\) weight_decay\u0026thinsp;=\u0026thinsp;0). The maximum number of epochs was 1000 and the batch size was 16. The early stopping threshold was set to 8. During the training process of each network, the early stopping threshold was hit after around 50 epochs. As we mentioned above, we trained one network with the AP images and another with the lateral images. We further combined the two models to see if jointly utilizing AP and lateral images could yield superior testing results. The combination was done via the following logistic-regression based method. We collected the outputs of the AP network and the lateral network (both are of the form of an 8-dimensional vector, with each element corresponding to the classification weight the network gives to the category of implants), and then fed them as the input to a two-layer feed-forward neural network, and trained the network with the data from the validation set. Note that the above construction relied on our dataset division procedure, where the training set, validation set, and testing set, each contained the AP and lateral images of the same set of patients. We referred to the resulting network constructed from the outputs of AP networks and Lateral network as the \u0026ldquo;joint network\u0026rdquo;.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\"\u003e\n \u003ch2\u003eModel Testing\u003c/h2\u003e\n \u003cp\u003eWe tested our models (AP, Lateral, Joint) using the testing set. The prediction result on each testing image was an 8-dimensional vector, with each coordinate representing the classification confidence of the corresponding category of implants.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\"\u003e\n \u003ch2\u003eStatistical Analysis\u003c/h2\u003e\n \u003cp\u003eSince we were studying a multi-class classification problem, we would directly present the confusion matrices of our methods on the testing data and compute the operation characteristics generalized for multi-class classification.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec12\"\u003e\n \u003ch2\u003eTraining Progress\u003c/h2\u003e\n \u003cp\u003eThe following figure (Fig.\u0026nbsp;3) showed the training progress of our method (we used the AP-network as an example). As the training proceeded, the network adjusted its parameters and learned the correct classification function as demonstrated by decrease in both training and validation loss.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\"\u003e\n \u003ch2\u003eTesting Performance\u003c/h2\u003e\n \u003cp\u003eIn this section, we evaluated our neural network models on the testing data sets. In the testing sets, we had in total 177 images. The classification results of the three neural network models (AP, Lateral, Joint) were presented in the form of tables (Table 1\u0026ndash;3). The column labels of each table indicated the ground-truth implant type, while row labels indicated the classification results. For example, the first row in the table of the AP-network indicated that the AP network classified 2 images with type-A plant as type A, and 1 image with type-A plant as type B.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\"\u003e\n \u003ch2\u003eAP-Network\u003c/h2\u003e\n \u003cp\u003eThe classification results (Confusion Matrix) of the AP-Network were shown in Table 1. The AP-network achieved an overall classification accuracy of 92.7%. Although the overall accuracy was high, the accuracy on some implants with scare training data was less satisfying. For example, its classification accuracy on implant type-G was about 58.3%.\u003c/p\u003e\n \u003cp\u003eTable 1: Classification Results (Confusion Matrix) of the AP-Network\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"832\" height=\"334\"\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\"\u003e\n \u003ch2\u003eLateral-Network\u003c/h2\u003e\n \u003cp\u003eThe classification results (Confusion Matrix) of the Lateral-Network were shown in Table 2. The Lateral network also achieved a high accuracy of 91.5%. It outperformed the AP-network on certain implant types (e.g., type G, with an accuracy of ~\u0026thinsp;91.7%) but underperformed on some other types (e.g., type E).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 2: Classification Results (Confusion Matrix) of the Lateral-Network\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAA0gAAAFLCAYAAADs7KxcAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAAABhaVRYdFNuaXBNZXRhZGF0YQAAAAAAeyJjbGlwUG9pbnRzIjpbeyJ4IjowLCJ5IjowfSx7IngiOjg0MCwieSI6MH0seyJ4Ijo4NDAsInkiOjMzMn0seyJ4IjowLCJ5IjozMzJ9XX3UMzfjAAA34klEQVR4Xu3dD2xd5X3/8Se/IU0aGp2dVmppEmW+2WALI8jYRkWBtRTiKFTpgLR2GJWYQE1sRLeWNAl2GN0aQRygErTENhoaUqntpKyFCGeJyZq2ztIkUMuhQYUp14sYayvVSRqqTZrW6f7u58lz0pObe+3ra3+dc67fL+nk/LVz/DnPPfd+7/k3L5fnAAAAAADu/4U+AAAAAMx5FEgAAAAAEFAgAQAAAEBAgQQAAAAAQcmbNMybNy8MAQAAAED1ipdEHEECAAAAgGDSI0iPPvqoO3PmjHvwwQfdkiVL/DTMPWoPJZoKpolsbZGvHbK1Q7Z2yNYO2dohWzvKVuL5TnoE6W/+5m/c5Zdf7q677jp3//33uzfffDPMAQAAAIDqMmmB9Ad/8Afusccecz/72c/c4sWL3W233ebWrl3rfvSjH4UlAAAAAKA6lH0N0hVXXOG2bNni/vM//9M1NTW5u+++23360592+/fvD0sAAAAAQLpNeg3SROc79vb2uq9//evuwx/+sPvCF77gVq9eHeag2nDuqx2ytUW+dsjWDtnaIVs7ZGuHbO0oW4nnO60CKfLiiy+6Z555xv3O7/yOL5R0Ch6qCy9MO2Rri3ztkK0dsrVDtnbI1g7Z2lG2Es93RgqkyHe/+11fKJ0+fdoXSvfdd1+Yg7TjhWmHbG2Rrx2ytUO2dsjWDtnaIVs7ylbi+c5ogRTZt2+fP/Xu7bff9oWSOqQbL0w7ZGuLfO2QrR2ytUO2dsjWDtnaUbYSz9fkQbHNzc3u1VdfdS+88II7ePCgu/LKK93jjz/u/uu//issAQAAAADJY1IgRZYvX+527drli6V/+7d/84XSI4884n75y1+GJS69gYEBXzkW69rb28NS0zM2Nlb0969cuTIsMbP0QF8VpgAAAACmxrRAitTX17t//Md/dEePHvXXJ33kIx9xX/rSl9zJkyfDEpfOu+++64aHh/1hta6uLtfW1uaHs9lsWGJqihVVdXV1/ndmMpnz/5d+v05FnKkiLKKiK77uKs62b98exgAAAABMZFYKpMhVV13lnn32WV8YXXbZZe7qq69269atc2+99VZYYvZt3LjRH+kqpKJmx44dYaw8Kka6u7vD2MT0+1WM6Wdm0t69e30hFnnyySfDEAAAAIDJzGqBFFmwYIE/qvGzn/3Mn3Z38803u8997nPu9ddfD0ski05Xi06Lix/xiU6T09/ygx/84Hxhouk6dW8i+p0qpu69994w5RydHhf9XxH9nxrXz+j/isb1f0Trpp8rpOX0f2zatOn8fPXVRT8LAAAA4LcuSYEUqa2tdY8++qgvlK699lp35513+u773/9+WOLS0xGem266yZ8Wp25oaMgXFypUdLrcli1bfAGycOHC86e2abnW1lY/XEi/SwWNCqPC5VRsbd261U/X6X4aj44waVpUTOnIlm6EITr61d/f74cLaTkdpdLvOnHihF9v/X79DTyrCgAAALjYJS2QIr/7u7/rvvzlL7v/+I//cLfeeqtbv369Lw4GBwfDEpeOrpsSFTXqVATpuqUbb7zRFzuHDh3yxYtOmSuHiip1xa5x0jVJKlz0/6joUlGj36uCRkd9NK5TAivV1NTkf79Ou9M6Fzu1EAAAAJjLElEgxem0MD0/SUdW/vZv/9YXId/+9rfD3EtDR2tUUERddN2ShkUFzVROV9PPRkeICqlwiv4fFUSivook/T/xU/ymKrpZhE5p1O+a7DRAAAAAYK5JXIEU0elkP/7xj/1DZp955hl/Jzw9V2m26aiLjuxExUR0Vzh1mqZiScWOjiRNRXQkKF7wqBCLj2tY/58KKRU3Kp5UKEV0JEt++MMf+nmTFU/RqYEqPnVann4OAAAAQEyuBM2aYPasGxwczOULiNxVV12Ve/bZZ8PUmaPfHf3NmUwmTD1neHj4/Lwok3xRdMHPRKJxzY/ki5fz0+Pz4tOjafq/o2n5IsYvE/9/NE3Uj6ZpfrTO8WX1s/q90bh+pnB+ubQ8bJCtLfK1Q7Z2yNYO2dohWztka0fZFuY7T//kJ15Ep2BJidmXjG7goCNKuuOdji6p0zVMsKX2kLS2UC3I1hb52iFbO2Rrh2ztkK0dsrWjbCWeb2JPsSvl4x//uPvOd77juzfffNPfJvwrX/mKfwAtAAAAAExH6gqkSGNjo/vmN7/pr6P5xS9+4QslXdfz3nvvhSUAAAAAYGpSWyBFli5d6np6evyd737zm9+4P/zDP3QPPPCAe+edd8ISAAAAAFCe1BdIkcWLF7uvfe1r/qGzegCt7j6nO+GNjIyEJQAAAABgYlVTIEU+9KEPua9+9au+ULrqqqvcpz71KffZz352Ss8pAgAAADA3VV2BFLn88svdww8/7AslPZhVR5Nuv/12/0wjAAAAACgmdbf5no7nn3/e3yJcp+A9+OCD7s477wxzMBluL2mHbG2Rrx2ytUO2dsjWDtnaIVs7ylbi+VbtEaRi7rvvPnfs2DH3+c9/3j355JP+OqUXX3wxzEW1WblypTt16lQYQyWUoXYcpTrN7+3tJecZMDAw4FpbW89nu2TJEtfe3u7Gxsbc9u3bw1KYjE6njrfRwk650m6nZ7L9QtQpY0yNXuvFsizs2CdUTq/7zs5OfzfkKE/tFzRN+1vti1GeUvvbqH1ONj/R8tVSUZo1weyq8Morr+Q++clP5v70T/8019PTE6aimLS1heHhYb/OXV1dYUpyJT3bbDabq6mp8eupXCOjo6O5lpYWP72hoSE3Pj4e5iRL0vNVbspPGau9Ku9o+uDgoJ9XmH1SJDnbjo4Ov37Nzc1hyjnKVNM0L5PJ+HacRElvt8pN61hqPZV/Uve/Sc9Wr/2ojRa+7rV/aGtrI9sK6fWvfa32qxqO3rfU7+/vP/9el0RJzlbtUeunfWqUadxk8y81rVthviXTLrZwtdq/f39u9erVucWLF+eeeuqp3P/+7/+GOYikrS3oDUTrrJ1dEl+McWnIttSbtUTzkvolQ5LzVduMCqBSH9SjZSiQpkZ5af0KC6RI0vcRSc42onWcaD35EF+56ANlsde92iv726mL9gnan5Z6zWs/rH1CEqUhW30xUsxk++NLTetWmO+cOsWulE9+8pPulVdecX19fe7111/3D53dunWre//998MSSBMdIh8aGnL5D0DuzJkz7qWXXgpzYOGWW27x/bNnz/o+yqfrIt944w3fVpctWxamXmj+/Plu586dYQwzZceOHS7/Zu33EY888kiYipmkh7dj5mmfsG7dujCGcui0Ot2sS7Q/VYbFaD/8+OOPhzFM1Qc+8IEwlH4USDEf+9jHXH9/v9u/f787efKkL5R0J7yf//znYQmkga4vU4G7YcMGP/7EE0/4PmzotSLVtGOcLdu2bfP9u+++2/dLqaur83fjxMz6whe+4PtcczCz9CUVj9awsWfPnjCEqXjttddcNpt1LS0tfn86EYpPCAVSEddee637h3/4Bzc6Our++7//2xdKf/3Xf+1OnDgRlkBSRRdY3nbbbX4nqG+ItVPkA9DMi24e0N3d7RoaGtyaNWvCHJRDN4zR0Quh+Lk0rr76at/XdlB7xszQkVHMPLXRV199NYxhKl5++WXfr6+v931gMhRIE9BdTZ5++mn/LKXf+73fc9ddd527//773ZtvvhmWQNLojXnz5s3nD59v2bLF95966infx/TcdNNN5+9Ck8lk3KZNm3xxNNEpCyju17/+dRjCpRL/Jln7eVQmfncqdY899liYg+kq3OdSyFfmV7/6VRgqrtjdA3W3RkyNPhMU5qhO7ThtKJDK8JGPfMSfk6o30MWLF7tbb73VrV271v3oRz8KSyAJdI6xjmbodu4RfTOvNxVd58EpH9M3PDysqxh9pyOsXV1d/gidiiSO0iFt4h82daYAKhPtE6JucHAwzMF0xfe54+Pj7o477ghzUIlS18rqejnlq88L0tPT4/bu3euHUT59JojvC6JO7ThtKJCm4IorrvBHJFQo6RlKum7g05/+tPuXf/mXsAQuJR090qkyH/zgBy/45kIf4EXXJWHm6GJWvanoYnflrmf28FyZ8v3Jn/xJGLrwgzpmz9tvv+37+lA02XUJKN+qVas4bdQAN2eo3PXXX+/7uoFTKcpXZw7J0qVLfR9zFwVSBS677DL3xS9+0f37v/+7fyPQhb46qrR79+6wBGabPpjrgvf4t21RpwKppqbG7du3z1/3gZmlh5uKiqSf/vSnfhiT05ux7l4n3Gnx0njmmWd8Xw8PB1C9ojNLOJsE5aJAmiZ9m/PWW2/520fqCMWNN97o74SH2aWjR/qgXuxbS30zrOuShNt3zrz4mw2nKU2N7rSo4l3F/UTFuzLmTX1m6YinvjTR6aHx03Ixc9RmuesakkCfA3TanKxevZqj9kaq6XEfFEgz5J577nFHjx71H3iee+45f/oRd/KZHXoT1gfMUs+RkWuuucb3dTOB3t5eP4zy6I1EbbuQjtrpuiO92YjOPeY0palRXgcOHPDDn/jEJ/yFwvFCScOdnZ3u0KFDnLI0Rf/8z/8chi6kD+y6+Dq6+6KuM+AGI1MXL9iLnVqrfYO+ONRZFpga5fm9733PD+v5jJgZ+kJbRZLOdtBrX/vbqFBS5to3cLfiykT7W53CWGx/oPcwUb7F5idSrgTNmmA2JpF/083l3xhymUwm9/TTT4ep6ZXUthA9nTnqij2lOXoi+WTLXSpan6RSToXZxTs9cbylpSU3ODgYfiJ5tJ5Jp6e6q53m37QvyLatra3ok/STIonZFu4TCjvtk9Vm+/v7w08kk9Y1qSbbL0RdqafqX2pat6Qq9n6lLsn7gTita9Jls1nfNuP7W3Ua1z43qe9nScy21P5W7bic+UkRrVfcPP2Tn3gRXdwuJWajTPqWTee5q//ggw/665Uuv/zyMDc91B5oCzbI1hb52iFbO2Rrh2ztkK0dsrWjbCWeL6fYGdNpMbt27fIPd3vnnXf8NRqPPPKI++UvfxmWAAAAAJAUFEizRE9vfuGFF/y1HKdPn/aF0pe+9CV38uTJsAQAAACAS40CaZZdddVV7tlnn/W3CNftwq+++mq3fv16fyc8AAAAAJcWBdIlsmDBAn8HFT109sMf/rC7+eab3ec+9zn3+uuvhyUAAAAAzDYKpEustrbWfeUrX/GF0rXXXuvuvPNO333/+98PSwAAAACYLdzFLoF27Njh73y3ePFif9e7JDxHgrun2CFbW+Rrh2ztkK0dsrVDtnbI1o6ylXi+FEgJpps6qFDSbcFVKH3mM58Jc2YfL0w7ZGuLfO2QrR2ytUO2dsjWDtnaUbYSz3dWT7HT08u1Euo0rCcY62nbl8KSJUuK/t/R+hXrpkN/61R/h55CPjIy4osjFUrXX3+9L5oAAAAA2Ji1AkkFiag6U7dlyxaXyWTcu+++66fPpvb2dpfNZsPYb6mIaWtr8+sXzY/Wt7m52Y9PhR4OGxVh+lsrpSNHw8PD7u///u/979Od73QaHgAAAICZNSsFku7WJnv37vV90QNUixUps0HFRbGCpa6urmThEV/3cm3dujUMuRn5W2+//Xa/Hj09PW7//v1u4cKF7oknnnD/8z//E5YAAAAAMB2zUiA999xzbsWKFWHst1SQbNy40Q9Hp6Dp6I76OvoSTYs6jUfToiNS0Wl70ZEajWs4mh4VZ6JhTdO8Smn9onX88z//85LrouF9+/a5tWvXXvD/6e/SMur0t0i0Xpqn3z2Zj3/84+473/mO7958803/0Nm/+7u/8w+gBQAAAFC5WSmQdPREd2SLi4oJdRqOjujoeUA6pU1HmFRURae49ff3+2VUVOl0s4iOqEQ/q+IiKkp0Cp9+5nvf+56fp6JFw/pdmlfJER0VMt3d3W5oaMj/nh/84Acl10XDOi1P6xA/+qSjSvpZzTt69KifFq1X/IhTORobG903v/lN98Mf/tDfJlyFkgrO9957LywBAAAAYCpmpUBS0XDy5Mkwdo6Khq6uLl8oaDgqWFpbW31fR1OiIzOi6VpW00vR6XFRUaICS06cOOH7urmBbnogmhcVMlOh4kPXKH3+858PU6YuXizFr79Soah5lVxbtHTpUtfb2+vefvtt95vf/MYXow888IB75513whIAAAAAyjErBZKOBOnIS3RKWbmi4qbaqTDSEaToaFqlVBh97Wtfcz//+c/9A2ibmprcX/3VX/k74QEAAACY3KwUSNFRkakctYlu4hBdWyQqmKIjQ9ERJxVdGtZpdRMVYDo1T6fWSbk/U66prkuh+B3+dIrgdH3oQx9yX/3qV/1pd3/8x3/sPvWpT7nPfvazEx59AwAAAJCX/1BelGZNMLsizc3N539v1A0PD/t50Xi+iPLjki82Llg2TstFy6vr7+/PtbW1nV+2q6vr/LD+31I/U4ymRz+rLlrH+O+M/2yp3xutTzQ/Go6vp5aN56L/w8LTTz/t/+/bb789t3fv3jC1fFo32CBbW+Rrh2ztkK0dsrVDtnbI1o6yLcx3nv7JT7yITveSErORUs8//7x/6KxOwdMDaO+4444wZ2JqD7QFG2Rri3ztkK0dsrVDtnbI1g7Z2lG2Es+XAmmO6u/v94XS//3f//lC6Z577glziuOFaYdsbZGvHbK1Q7Z2yNYO2dohWzvKVuL5UiDNcbt37/aF0i9+8Qv34IMPunXr1oU5F+KFaYdsbZGvHbK1Q7Z2yNYO2dohWzvKVuL5UiDB279/v/v617/uHzyrI0oqli677LIwlxemJbK1Rb52yNYO2dohWztka4ds7Shbiec7K3exQ/Ldeuut7pVXXnF9fX3+AbYf/ehH/YNr33///bAEAAAAUP0okHCBj33sY/76pNdee80/3PfKK690HR0dYS4AAABQ3SY9xQ4AAAAAqlm8JOIaJJSFc1/tkK0t8rVDtnbI1g7Z2iFbO2RrR9lKPF9OsQMAAACAgAIJAAAAAAIKJAAAAAAIKJAAAAAAIKBAAgAAAICAAgkAAAAAAgokAAAAAAgokAAAAAAgoEACAAAAgIACaRKnTp1yK1euDGNIg+3bt7va2lr/ZOTGxkZ38ODBMAfTNTY25lpbW322ylhZY2aQrR2ytUO2dsjWDtnaqZZsKZAm8fzzz7t9+/bxITslOjs7fX/37t2uv7/fZbNZt3r1av+CxfToy4KGhga/wxsfH3cHDhxw27ZtO585Kke2dsjWDtnaIVs7ZGunqrLNlaBZE8yeMzKZjM+hubk5TJmb0tAW8i/GXL4oCmPnaFzr3tPTE6YkT1peZ21tbbmamhqfcyTKd3h4OExJnjTkS7Z2yNYO2dohWztkayfN2RbmyxGkCQwMDLgVK1a4fJHkjyIdO3YszEESzZ8/3x/WjVuwYIHvL1y40PdRGX0r1N3d7V8PyjnS1NTk+319fb6PqSNbO2Rrh2ztkK0dsrVTbdlSIE1gy5YtbsOGDW7r1q1+vLe31/eRHocOHXLNzc1u1apVYQoqceTIEd+vr6/3/UhdXZ3vDw0N+T6mjmztkK0dsrVDtnbI1k61ZUuBVIKOHi1ZssRv2Ntuu83V1NT4yphrWdJB26m9vd2f+/qtb30rTEWljh8/HoaK07VeqAzZ2iFbO2Rrh2ztkK2dasuWAqmEp556yh9BEh0q3Lx5sx/WTRuQbLqhhk6LVEF75swZd8MNN3B6JAAAAMpCgVREdMe65cuX+76sWbPG9/WhW+dZIrm03XK5nBsdHXUtLS3+W4tNmzaFuQAAAEBpFEhF6JqjN954w9/DPep0REJ0RIKjSOmwbNkyf6qkrkHSTTZQuUWLFoWh4nRbT1SGbO2QrR2ytUO2dsjWTrVlS4FUQEePjh496u/frqMQ8W5wcNAvo+takB733ntvGEKlorvQjIyM+H4kOnVRD+RFZcjWDtnaIVs7ZGuHbO1UW7YUSAV09GjHjh0X3KIwojuh6WiEjiJxR7v0+MlPfuK3Gyqnm5W0tbX5u9DETzE9fPiw7+tuj6gM2dohWztka4ds7ZCtnarLNleCZk0wuyrpYaL6m0dHR8OUi3V0dPhl9CCsJD/0aqaloS0MDg767dLV1XX+IWV6QFlDQ0Mum8368SRKy+tMGSrf/A7Qj6v9a1yviSRLQ75ka4ds7ZCtHbK1Q7Z20pxtYb4l0y62cDXTh+rob1an8ULNzc0XLFNquWqkvzXp9MJUMRRtGw3Hi6WkSkO2EX15EL0OMpmM/1Ih6dKSL9naIVs7ZGuHbO2QrZ20ZluY7zz9k594Ed2YQErMxhyj9kBbsEG2tsjXDtnaIVs7ZGuHbO2QrR1lK/F8uQYJAAAAAAIKJAAAAAAIKJAAAAAAIKBAAgAAAICAAgkAAAAAAgokAAAAAAgokAAAAAAgoEACAAAAgGDSB8UCAAAAQDWLl0STFkglZmOO4QnOdsjWFvnaIVs7ZGuHbO2QrR2ytaNsJZ4vp9gBAAAAQECBBAAAAAABBRIAAAAABBRIAAAAABBQIAEAAABAQIEEAAAAAAEFEgAAAAAEFEgAAAAAEFAgAQAAAEBAgRSjJ+mW6mpra11ra6sbGBgISyOpxsbG/LaKttv27dvDHEzFnj173JIlSybNsbe39/xy6pS9tgGmhnZrh2xtHTt2zLW3t7uVK1eGKZgpZGuDfYKdasmWAikml8u5wcHBMHZuPOpefPFFl81m3dq1a/3OCsl06tQp19DQ4F+U4+Pj7sCBA27btm2us7MzLIFy6IuAe+65x73wwgu+/e/evdvnWNj2VRw9/PDD55dT5sp+xYoVflugPLRbO2RrS1+k7Nq1y3V3d4cpmClka4N9gp2qyjb/oaYozZpgdlUr9bfnN3aupqbGzxsdHQ1T54a0tIW2tja/jbStIv39/X79h4eHw5RkSVq2yk7r1NPTE6ac09XVdVGO+R1hrqOjI4ydk81mE5V30vItJo3tVsjWThqyjdP6Njc3h7FkI1s7aciWfYKdNGdbmC9HkKZg/vz5rqmpyQ//+te/9n0kh7650DdtOnqhbRWJtllfX5/vY2Kvvfaa7y9dutT3I2vWrPH9b3zjG74f+fGPfxyGzoleG1deeaXvY2K0WztkCyCOfYKdasuWAmmKTpw44fu///u/7/tIjiNHjvh+fX2970fq6up8f2hoyPcxsXfffdf333//fd+PRDnqVNPI/fff7/bt2+fPN9bOUd2mTZtcf3//+eUxMdqtHbIFEMc+wU61ZUuBVCZdKKkPgfpw2NbW5pYtWxbmICmOHz8ehoqLf7BHaddcc43vv/rqq75fKJ7junXr/Oth586d7oYbbnAPPPCA27Jli3+toDy0WztkCyCOfYKdasuWAmkCugNH1F133XX+Q2BLS4vbsWNHWAKoPqtWrfIXWepQuW7CILorTXQnmuhweUSvh+bmZr/z02vkrbfeCnMAAADShwJpArncb+9iNzw8fP6bct3SWEeUgGq1d+9e/2XA+vXr/d1onnzySXf27Fk/7/rrr/f9iO5sd++997rR0VFfWOlnuNMjAABIKwqkMi1fvtx/U64iSd+U6zoLJMuiRYvCUHH68I7y6AJL3epbXw6cPn3at319OSD33Xef70t0606dUqfTTlVYRUefDh486OdhYrRbO2QLII59gp1qy5YCaYruvvtu39eF6UiW6NSvkZER349ER/saGxt9H1On0+v0xUBXV9cFN1947LHH3OLFi8PYucJKRVJNTY07dOhQmIqJ0G7tkC2AOPYJdqotWwqkKYo+9GUyGd9HcuiDu47w6U4pupta5PDhw76/YcMG38fU6EiSjpgq240bN4ap56gQOnnyZBg7R0USr4/y0W7tkC2AOPYJdqou21wJmjXB7Ko1ODhY9G/PZrPnH5SpLskPvLKQlrag7aSHlOVfpH5c20njhQ8zTZKkZqvXQktLi1+/UvlFr4noobJ6OJyGlbm2RRKkoe2msd0K2dpJQ7YRPThd69vQ0HDBAyKTimztsE+wQ7Z2lG1hviXTLrZwtYv+5lJdJpPxG107rLlGf39aaPvoqePRNos+vCdV0rLVDk3rFO3kND4R5auc9TPqVFQlpTiSpOVbStrarZCtnbRkG+Ua7ybbZ1xqZGsnLdmyT7CT1mwL852nf/ITL6JbW0uJ2Zhj1B5oCzbI1hb52iFbO2Rrh2ztkK0dsrWjbCWeL9cgAQAAAEBAgQQAAAAAAQUSAAAAAAQUSAAAAAAQUCABAAAAQECBBAAAAAABBRIAAAAABJM+BwkAAAAAqlm8JOJBsSgLDyizQ7a2yNcO2dohWztka4ds7ZCtHWUr8Xw5xQ4AAAAAAgokAAAAAAgokAAAAAAgoEACAAAAgIACCQAAAAACCiQAAAAACCiQAAAAACCgQAIAAACAgAIJAAAAAAIKpBJOnTrlent7XWNjo3/CrjoNb9++3c9TH8k0NjbmWltb/Tarra1lW80gZalMle2SJUvcwMBAmIOZcOzYMdfe3u5WrlwZpmAmsE+wQ7Z2yNYO2dqplmwpkIrQh5Q/+qM/ck888YR76KGH3Pj4uMvlcm7nzp1u0aJFft6mTZvC0kgSFa8NDQ3+RantduDAAbdt2zbX2dkZlkCllOFzzz3nNm/e7Do6Otzp06fd2rVreWOZIXv27HG7du1y3d3dYQpmAvsEO2Rrh2ztkK2dqso2/8G/KM2aYHbVGh0dzdXU1OQymUwuv3HD1AtFy8wlaWkLbW1tftvEt11/f79f/+Hh4TAlWdKQbTabzbW0tISxc/Q60Lon/bWQtv2Y1re5uTmMJVsask3jPkHI1g7Z2iFbO2RrR+tXmC9HkAroyNCZM2fc1q1b3fz588PUCy1btszt2LEjjCEp9M2Fvn1fsWLFBduuqanJ9/v6+nwfU3f06FH32GOPhbFz9DrI7wz96wVIIvYJdsjWDtnaIVs71ZYtBVKMTq3bt2+fH9b5kxOZbD5m35EjR3y/vr7e9yN1dXW+PzQ05PuYOrX3KMe4xYsXu5qamjAGJAv7BDtka4ds7ZCtnWrLlgIp5vDhw77f3Nzs+0iX48ePh6HistlsGMJM+fa3v+2vSQKSiH2CHbK1Q7Z2yNZOtWVLgRRz9uzZMARgMjriqtPr7rvvvjAFAAAg/SiQAFRE1+v90z/90wXnGgMAAKQdBVLMNddc4/snTpzwfaSLbsE+Ed16EjNDz+rp6uryN2oAkop9gh2ytUO2dsjWTrVlS4EUs2rVKpfJZPx5kgcPHgxTkRbRnVJGRkZ8P6JTwUQP+sX06QHKN998M8UREo99gh2ytUO2dsjWTrVlS4FU4JlnnvH9L37xi/6WhaUMDAz4pwUjOXSnFN12WndKiW+76OYbGzZs8H1UTu3+Ax/4AHdxRCqwT7BDtnbI1g7Z2qm6bMPzkC6iWRPMrmrRQ630sFgNZ7PZMCeXGxwc9A/CUn8uSUtb0LbSQ8q0jUQPJtN4R0eHH0+itGSrDBsaGnJdXV0XdJpOvjMjeviuco4/aC+p0pBtGvcJQrZ2yNYO2dohWzvKtjDfkmkXW3gu0UbWBlaRFGWhYW1kzZtr0tQW9CGzubn5/Dbr6ekJc5IpDdmq3Uevg2KddoJJpfVLg6jNpiVXSUu2adsnCNnaIVs7ZGuHbO1oXQvznad/8hMvMm/ePN8vMRtzjNoDbcEG2doiXztka4ds7ZCtHbK1Q7Z2lK3E8+UaJAAAAAAIKJAAAAAAIKBAAgAAAICAAgkAAAAAAgokAAAAAAgokAAAAAAgoEACAAAAgGDS5yABAAAAQDWLl0Q8KBZl4QFldsjWFvnaIVs7ZGuHbO2QrR2ytaNsJZ4vp9gBAAAAQECBBAAAAAABBRIAAAAABBRIAAAAABBQIAEAAABAQIEEAAAAAAEFEgAAAAAEFEgAAAAAEFAgAQAAAEBAgRRs377dP0m3VNfY2OhaW1vdwMBA+Akk1djYmN9W2m61tbV+22LmHDt2zLW3t7uVK1eGKZgJtFtbarPxfbq6gwcPhrmoFO3WDtnaIVs71ZItBVKwceNGNz4+7pqbm/14V1eXy+VyvtP0hx56yL3xxhtu7dq1/oPhqVOn/HJIFm2XhoYG/6LUdjtw4IDbtm2b6+zsDEtgOvbs2eN27drluru7wxTMBNqtLb1hDw0N+f171LW1tbnly5eHJVAJ2q0dsrVDtnaqKtt8AVCUZk0wu2rlCyP/d6tfKL+xc/kN7+fn32DD1LkhLW0h/6EnV1NT47dVpL+/36//8PBwmJIsaXydpek1kIZ809huJS1tV/mOjo6GsXSg3dohWztka4ds7Wj9CvPlCNIUzJ8/3+3du9flN77bt28fp9sljL650JGNFStW+G0VaWpq8v2+vj7fB5KEdmtLp4Qq37vuusufZqejSZg+2q0dsrVDtnaqLVsKpCnSRs9XyH745Zdf9n0kw5EjR3y/vr7e9yN1dXW+r1NsgKSh3drSl1mZTMZls1n/5q3h3t7eMBeVot3aIVs7ZGun2rKlQKrAn/3Zn/m+rklCchw/fjwMFacPSEDS0G5t6frSEydO+PPhdW2pzgBYv349ZwBME+3WDtnaIVs71ZYtBVIFFixY4Pu8kAAgHXT0X8WSLhpWkaTT7QAAKIYCqQLvvfee7+tUDQBAeixbtsw9/vjj7syZM9zmGwBQFAVSBX7yk5/4vi5EQ3IsWrQoDBWnW08CSUO7nX1r1qwJQ6gU7dYO2dohWzvVli0F0hRFd+mQu+++2/eRDNGdUkZGRnw/ortYiR72CyQN7fbS0Gl2PAepcrRbO2Rrh2ztVFu2FEhToOJID4nVqRm6kx1vrsmiO6Vou+hOKdpWkcOHD/v+hg0bfB9IEtrt7HvppZfc5s2bwxgqQbu1Q7Z2yNZO1WUbnod0Ec2aYHZV0oOt9PBL/d3xB8Vquh50lclk/Lx8Awhz5o60tIVsNusfUhZtIz2YTOMdHR1+PInS9jrTAze1zg0NDRc8DC6p0pBvGtutJD3bqK0qx6it9vT0pGIfTru1Q7Z2yNYO2dpRtoX5lky72MLVTAVR9DcX6/RhUBtcG3suUgZpoQ9FUaGrolYfiJIsTdlGuca7pL8m0pJv2tqtJD3b+Jde6jSsL7vSgHZrh2ztkK0dsrWjdS3Md57+yU+8yLx583y/xGzMMWoPtAUbZGuLfO2QrR2ytUO2dsjWDtnaUbYSz5drkAAAAAAgoEACAAAAgIACCQAAAAACCiQAAAAACCiQAAAAACCgQAIAAACAgAIJAAAAAIJJn4MEAAAAANUsXhLxoFiUhQeU2SFbW+Rrh2ztkK0dsrVDtnbI1o6ylXi+nGIHAAAAAAEFEgAAAAAEFEgAAAAAEFAgAQAAAEBAgQQAAAAAAQUSAAAAAAQUSAAAAAAQUCABAAAAQECBBAAAAAABBVLMwYMH/dN0J+tqa2vDTyCJxsbGXGtr6/lttX379jAHM6G9vf2i14ReO5ge2q0dsrXFPsEG7dYO2dqplmwpkGKWL1/ucrmca2lp8eNdXV1+PN6Njo76DY5kOnXqlGtoaPDbaHx83B04cMBt27bNdXZ2hiUwHdrxDQ0Nuebm5vNdW1ubf+2gcrRbO2Rri32CDdqtHbK1U1XZ5j/0F6VZE8yuavnCyP/t6hdTano1S0tbyL8x52pqanL5F2aYksv19/f79R8eHg5TkiVNrzPlOzo6GsbSIQ35prHdCtnaSct+gX2CDdqtHbK1k+ZsC/PlCFIFNm7cGIaQJPrmoru7261YscLNnz8/THWuqanJ9/v6+nwflTl27JjP96677vKn1OibY0wf7dYO2dpin2CDdmuHbO1UW7YUSFM0MDAQhpA0R44c8f36+nrfj9TV1fm+TgNB5fbt2+cymYzLZrN+J6jh3t7eMBeVot3aIVtb7BNs0G7tkK2dasuWAmkK9O3Yyy+/HMaQNMePHw9DxelNHJXTkdMTJ07484p1fV5NTY1bv349XxpME+3WDtnaYp9gg3Zrh2ztVFu2FEgT2LRp0/k78qjTt2O/+tWvwlxgbtKhc30w0sWX+kCkU2sAzF3sEwBUGwqkCegbsVzut3ew07dkt9xyS5gLzG3Lli1zjz/+uDtz5gy39AXAPgFA1aBAmoLoWzIk06JFi8JQcbr1JGbWmjVrwhAqRbu1Q7azj33C9NFu7ZCtnWrLlgIJVSO6U8rIyIjvR3SnJWlsbPR9zCydUsMzTypHu7VDtpcG+4Tpod3aIVs71ZYtBVKFeKBY8uhOKW1tbf5OKbrdZOTw4cO+v2HDBt/HzHnppZfc5s2bwxgqQbu1Q7azj33C9NFu7ZCtnarLVg9DKkazJphdtfRwq5aWFv+3d3R0hKm/lc1m/fyenp4wZW5IS1vQ9qmpqfEPKxM9mEzjxbZlUqQhWz0IMnpNRA+A02sgyjnJ0pBvGtutkK2dpGfLPsEW7dYO2dpJc7aF+ZZMu9jC1U4bMvq7J+uiN4S5Qn9zWuiNu7m52a9zJpNJfDGbhmzV3qNM1WlYT8dOg7S03bS1WyFbO0nPln2CPdqtHbK1k9ZsC/Odp3/yEy+i21pLidmYY9QeaAs2yNYW+dohWztka4ds7ZCtHbK1o2wlni/XIAEAAABAQIEEAAAAAAEFEgAAAAAEFEgAAAAAEFAgAQAAAEBAgQQAAAAAAQUSAAAAAASTPgcJAAAAAKpZvCTiQbEoCw8os0O2tsjXDtnaIVs7ZGuHbO2QrR1lK/F8OcUOAAAAAAIKJAAAAAAIKJAAAAAAIKBAAgAAAICAAgkAAAAAAgokAAAAAAgokAAAAAAgoEACAAAAgIACCQAAAAACCqQCeppuOR2Sa2xszLW2tvrtVFtb67Zv3x7mYCYdO3bMZ3zw4MEwBdNBu7VDtnbIdmYoN+WnHBsbG0vuV8tdDhOj3dqplmwpkArkcjk3ODjoh5ubm/14vNO8mpoaPx/Jc+rUKdfQ0OBflOPj4+7AgQNu27ZtrrOzMyyBmaCc77rrrjCG6aLd2iFbO2Q7M6K8du/e7fr7+102m3WrV6/2HzTjyl0OE6Pd2qmqbPMf+ovSrAlmVz397fkCKYxdqKurKwzNHWlpC21tbbl8AZvLvzDDlFwu/0bi1394eDhMSZY0vs5aWlr86yDJuUbSkG8a262QrR2ytZOkbJWdMouLMuzp6QlTyl/uUqPd2iFbO1q/wnw5glSBjRs3hiEkib656O7uditWrHDz588PU51ramry/b6+Pt/H9PT29rr6+np34403himYDtqtHbK1Q7YzQ9npdKS4BQsW+P7ChQt9X8pdDhOj3dqptmwpkKZAG5/zfZPryJEjvq8P73F1dXW+PzQ05PuonK47+u53v8uXBDOIdmuHbO2QrZ1Dhw75U/xXrVoVphRX7nL4LdqtnWrLlgJpCp5//vkwhCQ6fvx4GCpO52ujcvqCYNOmTe5b3/pWmIKZQLu1Q7Z2yHbm6Vqi9vZ2f83GRPvZcpfDxWi3dqotWwqkCezbt8/fhSPq9OEQmKseeOAB19XVdcGhcwDA9OnslEwm409ROnPmjLvhhhv8EftC5S4HYHookCZQeBe74eHhMAeYW3Td0V/8xV+4ZcuWhSkAgJmyfPly/zljdHTUtbS0+G/bi30pW+5yAKaHAmkKtGNSh2RatGhRGCpOt55EZdavX+/Wrl17wRHVm266yc9TX+OoDO3WDtnaIVsb+hJqYGDAf0Grs1hKKXc5XIh2a6fasqVAQtWI7pQyMjLi+5Ho9AM9VA+V0al1hV1bW5ufp77GURnarR2ytUO2tu69994wNLFyl8M5tFs7VZdtrgTNmmB21dPf3lziOUhzUVraQrF78OsZEVr/bDYbpiRLWl9neqaB1j3JzzaQNOSbxnYrZGuHbO2kIduOjo6yPoOUu9xsod3aIVs7Wr/CfEumXWzhuWJwcPD835/kDTqb0tIWtL304tSLVPThXeN6E0mqtL7OKJBmThrbrZCtHbK1k6Rs9XlDmXV1dZ3/UKkHazY0NFzw+aPc5S412q0dsrWjbAvzLZl2sYXngujvjnccSUrHCzMyOjrqt5nWOZPJ+G8vkixN2cZpx6d1Vz/J0pJv2tqtkK0dsrWTpGz1gVJFjtZJnYbjRVCk3OUutSRlOxHarZ20ZluY7zz9k594keii6xKzMceoPdAWbJCtLfK1Q7Z2yNYO2dohWztka0fZSjxfbtIAAAAAAAEFEgAAAAAEFEgAAAAAEFAgAQAAAEBAgQQAAAAAAQUSAAAAAAQUSAAAAAAQUCABAAAAQDDpg2IBAAAAoJrFS6JJC6QSszHH8ARnO2Rri3ztkK0dsrVDtnbI1g7Z2lG2Es+XU+wAAAAAIKBAAgAAAICAAgkAAAAAAgokAAAAAAgokAAAAAAgoEACAAAAgIACCQAAAAACCiQAAAAACCiQAAAAACCgQCrh2LFjrr293S1ZssQ/YVddY2Oj6+3tdadOnXIDAwNhSSTN2NiYa21t9dustrbWbd++PczBTIheGytXrgxTMBNot3bI1g7Z2iHbmVPu+xbvb9NXLe2WAqkIbczrrrvOnT592r3wwgsul8v5bufOne7s2bPugx/8oJ+O5FHx2tDQ4F+U4+Pj7sCBA27btm2us7MzLIHp2LNnj9u1a5fr7u4OUzATaLd2yNYO2doh25lT7vsW72/TV1XtNv/BvyjNmmB21erq6vJ/d1tbW5hysf7+/lxzc3MYmxvS0ha03WpqanL5F2aYcm57af2Hh4fDlGRJ4+tM65yW10Aa8k1juxWytUO2dsjWTpKz1bqV875V7nKzjXZrR+tXmC9HkGJ0aHXTpk0uk8m4HTt2hKkX06HDO+64I4whKfTNhb75WbFihZs/f36Y6lxTU5Pv9/X1+T6QJLRbO2Rrh2ztkC3SqNraLQVSjK4vki9/+cu+P5F169aFISTFkSNHfL++vt73I3V1db4/NDTk+0CS0G7tkK0dsrVDtkijamu3FEgx0Y0Xli5d6vtIl+PHj4eh4rLZbBgCkoN2a4ds7ZCtHbJFGlVbu6VAijlz5kwYKk53NdFdOeIdd5UBAAAAqgcFUhHvv/9+GLrQ3r173ejoaBhzfnjjxo1hDAAAAEDaUSDF6NaE8q//+q++X8yyZcvC0IXDuPQWLVoUhoqLti+QJLRbO2Rrh2ztkC3SqNraLQVSzEMPPeT7uguH7saBdInulDIyMuL7Ed2dUPSgXyBpaLd2yNYO2dohW6RRtbVbCqQY3b67paXFX4v0l3/5lxRJKaM7pbS1tfk7pcS33eHDh31/w4YNvg8kCe3WDtnaIVs7ZIs0qrp2G56HdBHNmmB21dLDrfIb2P/tmUzGP+AqeuBVNps9/8CruZZNWv5ebSM9pEzbUPRgMo13dHT48SRKW1saHR3169zQ0HDBw+CSKg35prHdCtnaIVs7ZGsnqdmW+76V5Pc32q0dZVuYb8m0iy08l2ijagOrSIqyUKenK2tD60U0l6SpLWjbaDtpnbX9enp6wpxkSlO2Ua7xTq+VJEtLvmlrt0K2dsjWDtnaSWK25b5vlbvcpaL1SYO0ttvCfOfpn/zEi+gW1lJiNuYYtQfagg2ytUW+dsjWDtnaIVs7ZGuHbO0oW4nnyzVIAAAAABBQIAEAAABAQIEEAAAAAAEFEgAAAAAEFEgAAAAAEFAgAQAAAEBAgQQAAAAAwaTPQQIAAACAahYviXhQLMrCA8rskK0t8rVDtnbI1g7Z2iFbO2RrR9lKPF9OsQMAAACAgAIJAAAAAAIKJAAAAAAIKJAAAAAAIKBAAgAAAICAAgkAAAAAAgokAAAAAAgokAAAAAAgoEACAAAAgIACKdi+fbt/km5hd/DgwbLmIznGxsZca2ur3z61tbV+22FmkK0dsrVDtnbI1g7Z2iFbe6dOnXK9vb1u5cqVPmf104QCKdi4caMbHx93LS0tflz9XC7nli9f7scnm49k0AuyoaHB7/C0vQ4cOOC2bdvmOjs7wxKoFNnaIVs7ZGuHbO2QrR2ytXfs2DF3ww03uIcfftjdcsstbnR01O3duzfMTYn8h/yiNGuC2VWrq6vL/92Dg4NhyoWi+erPJWlpC21tbbmamppcfqcXpuRy/f39fv2Hh4fDlGQhW1tpyJds7ZCtHbK1Q7Z2yNZWvhjy+eaL0AsyTjJlW5gvR5BKuOKKK8IQ0kLfCnV3d7sVK1a4+fPnh6nONTU1+X5fX5/vY+rI1g7Z2iFbO2Rrh2ztkK0t5fuJT3zCH53TEaN4xmlDgYSqceTIEd+vr6/3/UhdXZ3vDw0N+T6mjmztkK0dsrVDtnbI1g7Z2nrkkUfcmTNn3NatW1NdHAkFEqrG8ePHw1Bx2Ww2DGGqyNYO2dohWztka4ds7ZCtHR09GhgYcDU1Ne7s2bNuyZIl/uYM6mt62lAglXDTTTf5DVvYbdq0KSwBAAAA4Kc//ak/epTJZNzChQvdiRMnfMGpgmnt2rVuz549Ycl0oEAqYXh4WFdrXdR1dXWFJQAAAAC89dZbvv+Zz3zGrVq1yg/r1MWdO3f64UcffdT304ICCVVj0aJFYag43dYTlSFbO2Rrh2ztkK0dsrVDtnZ01KgYFUnK9Y033ghT0oECCVUjugvNyMiI70d0P35pbGz0fUwd2dohWztka4ds7ZCtHbK189GPftT3T5486ftxumGDTrVLEwqkEt5///0whLTQtxRtbW3+LjS6WDBy+PBh39+wYYPvY+rI1g7Z2iFbO2Rrh2ztkK2dZcuWuebmZn9Dhni2cvToUZ97quRK0KwJZlclPdAqv3H9353fkGHqb2l+S0uLn6/+XJKWtpDNZv0DyqLtp4e+abyjo8OPJxHZ2kpDvmRrh2ztkK0dsrVDtnaih8QqW31mVqfhhoQ/NFbZFuZbMu1iC1ezrq6u839zvNMLp5z51U5/a1roBRoVuplMJtfT0xPmJBPZ2kpLvmRrh2ztkK0dsrVDtnbi2caLpSTTuhbmO0//5CdeRLe0lhKzMceoPdAWbJCtLfK1Q7Z2yNYO2dohWztka0fZSjxfrkECAAAAgIACCQAAAAACCiQAAAAACCiQAAAAACCgQAIAAACAgAIJAAAAAAIKJAAAAAAIJn0OEgAAAABUM56DBAAAAABFlDyCBAAAAABzDUeQAAAAAMBz7v8DzXMXTDF/jLQAAAAASUVORK5CYII=\" width=\"840\" height=\"331\"\u003e\u003cbr\u003e\u003c/p\u003e\n \u003ch2\u003eJoint Network\u003c/h2\u003e\n \u003cp\u003eThe classification results (Confusion Matrix) of the Joint Network were shown in Table 3. The join-network that essentially combined the AP-network and the Lateral-network had an overall accuracy of 95.4%. In addition to its higher overall accuracy, it also achieved great classification accuracy on each individual implant type. Essentially, from the experiments, we saw that combining AP images with lateral images had the potential to achieve the best of both worlds, and improved the classification performance of only using one category of image data.\u003c/p\u003e\n \u003cdiv\u003e\u003cstrong\u003eTable 3: Classification Results (Confusion Matrix) of the Joint Network\u003c/strong\u003e\u003c/div\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" style=\"width: 673px; height: 264.449px;\" width=\"673\" height=\"264.449\"\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\"\u003e\n \u003ch2\u003eOperation Characteristics\u003c/h2\u003e\n \u003cp\u003eFor a multi-class classification method, let \\(M\\) be its confusion matrix on the testing data. We used \\({M}_{ij}\\) to denote the entry on the \\(i\\)th row and \\(j\\)th column of \\(M\\). Then, for each row (class), the recall and precision of the algorithm were defined as:\u003c/p\u003e\n \u003cdiv id=\"Equa\"\u003e\n \u003cdiv id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$Precisio{n}_{i}=\\frac{{M}_{ii}}{{\\sum }_{j}{M}_{ji}}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Equb\"\u003e\n \u003cdiv id=\"FileID_Equb\" name=\"EquationSource\"\u003e$$Recal{l}_{i}=\\frac{{M}_{ii}}{{\\sum }_{j}{M}_{ij}}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eThe precision (recall) of the algorithm was the average of its precision (recall) for each class.\u003c/p\u003e\n \u003cp\u003eFollowing the above definition, we had the following operating characteristics of the AP-network, Lateral-network and Joint-network (Table 4).\u003c/p\u003e\n \u003cp\u003eTable 4 Operation Characteristics of Our Methods\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003eAP-network\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003e0.983\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003e0.802\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003eLateral-network\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003e0.961\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003e0.849\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003eJoint-network\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003e0.987\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"33.333333333333336%\" valign=\"top\"\u003e\n \u003cp\u003e0.887\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003ePreoperative identification of arthroplasty implants prior to revision surgery is a difficult task and an essential step in preventing increases in perioperative morbidity and increased healthcare costs. The key finding of this study was that first, convolutional neural networks could be trained to provide an automated identification of knee arthroplasty implant from radiographic images with very high accuracy. Second, training the neural networks with both AP and lateral images further improve the accuracy compared with training with one kind of images.\u003c/p\u003e \u003cp\u003eTo our knowledge, there are only 2 previous studies that have used deep learning for identifying knee arthroplasty implants from plain radiographs. One study by Yi et al.[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] used a small sample sized dataset (273 AP knee radiographs) to distinguish between TKA and UKA, and to distinguish between only two TKA models. Karnuta et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]reported in the other study that their deep learning model using 682 AP knee radiographs could differentiate successfully between 9 unique knee arthroplasty implants. However, they only evaluated AP knee radiographs, as opposed to clinical practice, where lateral radiographs should always be needed, especially in differentiating cruciate retaining vs. PCL-sacrificing implants. In our study, the deep learning model developed from over 1000 AP and lateral view of knee radiographs represents the first advanced artificial intelligence model in classifying knee arthroplasty implants from radiographs.\u003c/p\u003e \u003cp\u003eThis study has several limitations.\u003c/p\u003e \u003cp\u003eFirst, the data come from only one institution, the number of images for each brand implant was highly varied because the contractual relationship between the single institution and different manufactures has been changing over the past 10 years. This could result in an imbalance in trained implant data.\u003c/p\u003e \u003cp\u003eSecond, our algorithm was trained with only 8 knee arthroplasty implants, therefore, the data was limited to these models alone and should not be generalized to identify other models.\u003c/p\u003e \u003cp\u003eThird, deep-learning based methods require large-scale training data sets. Collecting and labeling such data sets requires the combined effort of a team involving a surgeon, radiologist, and computer scientist. Also, the trained CNN functions more like a black box and is not as humanly interpretable, which might be a drawback in clinical practice.\u003c/p\u003e \u003cp\u003eIn conclusion, the technology behind machine learning is exponentially advancing and the enthusiasm for AI in healthcare is growing. With this study, we have demonstrated that a deep learning algorithm can identify the design of 8 different knee arthroplasty implants from AP and lateral knee radiograph. These findings suggest that this technology has the potential to classify knee implants prior to revision surgery, thus saving significant time, and reducing perioperative morbidity and healthcare cost. It is hoped that it can be used to collect large-scale implant information and may be applied to mobile phone applications in the future.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eZ.G. and X.H. contributed equally to this work. Z.G.: Resources, Investigation, Data curation, Writing\u0026mdash;original draf. X.H.: Data curation, Project administration, Writing\u0026mdash;review \u0026amp; editing M.H.: Data curation, Resources, Writing\u0026mdash;review \u0026amp; editing X.F.: Software, Formal analysis, Writing\u0026mdash;review \u0026amp; editing.Y.F.: Conceptualization, Methodology, Writing\u0026mdash;review \u0026amp; editing, Supervision.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets used and/or analysed during the current study available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References ","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBozic KJ, Kamath AF, Ong K et al: Comparative Epidemiology of Revision Arthroplasty: Failed THA Poses Greater Clinical and Economic Burdens Than Failed TKA. Clin Orthop Relat Res 2015, 473(6):2131-2138.10.1007/s11999-014-4078-8\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilson NA, Jehn M, York S, Davis CM, 3rd: Revision total hip and knee arthroplasty implant identification: implications for use of Unique Device Identification 2012 AAHKS member survey results. J Arthroplasty 2014, 29(2):251-255.10.1016/j.arth.2013.06.027\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeCun Y, Bengio Y, Hinton G: Deep learning. Nature 2015, 521(7553):436-444.10.1038/nature14539\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEsteva A, Kuprel B, Novoa RA et al: Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017, 542(7639):115-118.10.1038/nature21056\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim DH, MacKinnon T: Artificial intelligence in fracture detection: transfer learning from deep convolutional neural networks. Clin Radiol 2018, 73(5):439-445.10.1016/j.crad.2017.11.015\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbedin J, Antony J, McGuinness K et al: Predicting knee osteoarthritis severity: comparative modeling based on patient's data and plain X-ray images. Sci Rep 2019, 9(1):5761.10.1038/s41598-019-42215-9\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S, Sun J: Identity mappings in deep residual networks. In: Computer Vision\u0026ndash;ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11\u0026ndash;14, 2016, Proceedings, Part IV 14: 2016: Springer; 2016: 630\u0026ndash;645\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrizhevsky A, Sutskever I, Hinton GE: ImageNet classification with deep convolutional neural networks. Communications of the ACM 2017, 60(6):84-90.10.1145/3065386\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma DP, Ba J: Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980 2014\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYi PH, Wei J, Kim TK et al: Automated detection \u0026amp; classification of knee arthroplasty using deep learning. Knee 2020, 27(2):535-542.10.1016/j.knee.2019.11.020\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKarnuta JM, Luu BC, Roth AL et al: Artificial Intelligence to Identify Arthroplasty Implants From Radiographs of the Knee. J Arthroplasty 2021, 36(3):935-940.10.1016/j.arth.2020.10.021\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":"Machine learning, Deep learning, Artificial intelligence, Knee arthroplasty, Neural networks","lastPublishedDoi":"10.21203/rs.3.rs-4299072/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4299072/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e A critical step in preoperative planning for revision knee arthroplasty is the identification of the failed implant. The purpose of this study was to develop and evaluate the performance of deep learning methods based on convolutional neural networks (CNN) to detect and identify specific knee arthroplasty models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eIn this study, we propose a novel deep learning-based approach to identify knee arthroplasty implants’ design using both anterior-posterior (AP) and lateral images. We harness the pre-trained ResNet50 CNN model and employ transfer learning methods to adapt the model for implants identification task using a total of 1814 radiographs of 8 different knee arthroplasty implant designs. Performance was compared with operative note and crosschecked with implant sheets. We also evaluate the difference of performance of models trained with AP images, lateral images or both.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eThe training and validation data sets were comprised of 907 AP and 907 lateral view radiographs across 733 patients and included 8 knee arthroplasty implants from 5 leading implant manufactures. After 1000 training epochs the model classified 8 implant models with very high accuracy. Our results showed that jointly using AP images and lateral images improved the classification accuracy of the CNN model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eCNN can accurately distinguish between specific knee arthroplasty designs. This technology could offer a useful adjunct to the surgeon in preoperative identification of the prior implant. Furthermore, using both AP images and lateral images to train the CNN is more effective than using images from only one perspective.\u003c/p\u003e","manuscriptTitle":"Deep Learning-based Automated Identification of Knee Arthroplasty Implants","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-05-10 21:34:20","doi":"10.21203/rs.3.rs-4299072/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":"64103f63-3346-42f2-81bd-36bb2a76ef23","owner":[],"postedDate":"May 10th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":31696000,"name":"Health sciences/Diseases"},{"id":31696001,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2024-06-18T08:15:53+00:00","versionOfRecord":[],"versionCreatedAt":"2024-05-10 21:34:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4299072","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4299072","identity":"rs-4299072","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}