Session 3: Deep Learning


  • Automated identification of orthopedic implants in radiographs using deep learning 

Ravi Patel 1,2 MB BChir MA (Cantab) ; Elizabeth H.E. Thong1 MBBS BSc. Vineet Batta3 FRCS(Orth), MBBS (Presenting Author). Anil Bharath2 PhD, BEng , Darrel Francis1 FRCP MD MRCP MB BChir. James Howard1 MRCP MB BChir

1.Faculty of Medicine, Imperial College Healthcare NHS Trust, London, UK. 2.Department of Bioengineering, Imperial College London, London, UK. 3.Department of Orthopaedic Surgery, Luton and Dunstable University Hospital, Luton, UK 


This study presents the development and evaluation of a neural network classifier for identifying the model of metallic joint implants from plain film radiographs, with network performance compared to human expert performance. 

Material and Methods: 537 knee and 1191 hip unilateral anterior-posterior radiographs of orthopedic implants, containing 12 implant models, were used to develop a range of neural network classifiers. A subset of 315 images paired with auto-generated image masks were used to develop a segmentation network to automatically crop implants from radiographs for subsequent classification. Networks taking original radiographs and cropped implants as inputs were ensembled together to provide a consensus opinion. Accuracies of three orthopedic experts assisted by a reference radiographic image gallery were evaluated for performance comparison. 

Results: A consensus prediction from a combination of neural networks performed best, by processing artificial intelligence-cropped implants as well as the full radiographs, and achieving a 99.4% accuracy and 100% top 3 accuracy. The network performed superiorly to the median and best human experts (64.4% and 76.7% accuracy, respectively). The network is robust to scan quality variation and difficult to distinguish implants, and predictions are 100% reproducible. 

Interpretation and conclusion: This study demonstrates, for the first time, supra-human expert performance of a neural network classifier at identifying orthopedic implant models in radiographs. Its real-world application can be readily realized through training on a broader range of implants encompassing all joints, and across institutions. All code has been made freely available to support this endeavor. Current systems are inadequate, associated with poorer patient surgical outcomes, significant time burdens on clinical staff, and increased healthcare costs. 

Keywords: Automated identification, Orthopedic implants, Deep learning 


  • Unsupervised Machine Learning to Identify Implant Migration and Patient Demographic Profiles in Total Knee Arthroplasty

Elise Laende 1,2, Joshua Britton 1, 2, Kathryn Young-Shand2, Glen Richardson2, Michael Dunbar2. 1Queen’s University, Kingston, ON, 2.Dalhousie University, Halifax, NS.

Introduction: Unsupervised machine learning techniques such as cluster analysis provide the opportunity to investigate relationships between demographic factors and tibial component migration following total knee arthroplasty (TKA). The objective of this study was to identify distinct clusters of TKA subjects based on implant fixation, demographic factors, and tibial component migration.

Materials and method A dataset of implant migration (maximum total point motion (MTPM) at 1 year and change from 1 to 2 years), implant characteristics, and subject demographics was used. The cluster algorithm and optimal number of clusters were selected from exploratory analyses based on maximizing the average silhouette coefficient for clusters k=2:6.  Subjects were stratified into 4 clustering datasets by sex and fixation to permit numerical methods. Hierarchical agglomerative cluster methodologies with complete linkages were applied to each group using features of age, body mass index (BMI), and implant migration measures.

Results 360 TKA were included: 222 cemented (68% female) and 138 uncemented (49% female).  Outlier analysis using Tukey’s outer fence removed 19 TKA.  Two clusters (k=2) yielded maximum silhouette coefficients.  Clusters for both the female groups (cemented and uncemented) were separated based on age and BMI, while both male groups had clusters based on BMI and implant migration. The young, high BMI cluster in the female uncemented group had acceptable implant migration. The high BMI cluster in the cemented male group had lower 1-year migration, while the high BMI cluster in the uncemented male group had higher 1-year migration higher migration, but acceptable 1 to 2 year migration.

Interpretation and conclusion TKA patients are not a homogenous group and disaggregating data by sex prior to analysis should be considered.  These findings indicate it would be appropriate for most female patients to receive uncemented implants, including those who may be considered a non-ideal candidate for TKA (young, high BMI).

Keywords: machine learning, cluster analysis, total knee arthroplasty, implant migration


  • Deep Learning-Based Reconstruction for Sparse-View Cone-Beam Computed Tomography to Assess Implant Migration

Rudy Baronette1,2,3, Xunhua Yuan2, Matthew G Teeter1,2,3,4,5, David W Holdsworth1,2,3,4

1Western Bone and Joint Institute. 2Imaging Research Laboratories, Robarts Research Institute

Depts. of Medical Biophysics3 and Surgery,4 Western University. 5Lawson Health Research Institute, London ON, Canada

Radiostereometric analysis (RSA) and flat-panel cone-beam CT (CBCT) scanners share fundamental imaging principles used to evaluate orthopaedic implant micromotion, polyethylene wear, and kinematic analysis. Routine clinical adoption of RSA was limited by the requirement of two x-ray sources, two x-ray detectors, and a calibration cage. Recent advances in deep learning allowed for sparse-view reconstruction; however, these studies are limited to the head and torso. To address this limitation, we propose a new deep learning-based approach which was trained and evaluated with a knee phantom to explore applications to RSA.

We trained a deep convolutional network that combines sparse-view FDK reconstruction with a multiresolution convolutional neural network (CNN), based on U-Net architecture. Our training dataset consisted of 6 3D volumes acquired using a ceiling-mounted x-ray system equipped with a flat-panel detector that has a 43 x 35 cm field of view (DRX3543C, Carestream, USA). The FDK algorithm was used to generate volumetric reconstructions from 60- and 217-views. The slices from sparse-view reconstructions were used as an input to the network, while slices from the 217-view reconstruction were used as the target image. The 3D images were split into individual slices resulting in a dataset of2D images used to train a CNN, modified with 20% dropout layers after each convolution to increase regularization and reduce overfitting. Performance was evaluated using mean absolute error (MAE) and peak signal-to-noise ratio (PSNR).

On the test dataset of 227 2D images, our proposed method performed with 0.084 MAE and 22.16 PSNR. Our proposed method similar performance on both metrics when compared to previously published deep learning algorithms.1

We present a deep learning-based approach to reduce limited angle artifacts and improve image quality for sparse-view CBCT volumes reconstructed with the FDK algorithm. This method has potential to allow CBCT to supplant RSA in some clinical applications.

Keywords: sparse-view cone-beam CT, radiostereometric analysis, deep learning


  • Accuracy of an Autonomous Method for Extracting 3D Knee Replacement Kinematics from Dynamic Single Plane Fluoroscopic Images

Andrew Jensen, Paris Flood, Lindsey Palm, Scott Banks

Introduction: There are currently no inexpensive, reliable methods for clinically analyzing dynamic joint kinematics. Model-image registration using single- or bi-plane radiographic images proves to be effective for analyzing joint kinematics, but required human supervision is too cumbersome for practical clinical use. With recent machine learning (ML) advancements, the need for human supervision might be eliminated, bringing objective dynamic joint kinematics assessment into practical clinical use.

Purpose/Aim of Study: Our study aims to determine the efficacy of ML methods to measure dynamic 3D knee replacement kinematics autonomously (i.e. without human supervision). We will quantify the algorithm’s performance and determine if these methods produce results that are comparable to human-supervised kinematics measurements.

Materials and Methods: We trained 4 convolutional neural networks (NNs) to segment and estimate the pose of total knee arthroplasty implants in single-plane fluoroscopic images. The first 2 NNs segmented the femoral and tibial components in each image. We trained the networks using 2982 images from a single research study where human-supervised kinematic results were available. Two additional NNs are for direct pose-estimation, with the femoral and tibial models trained on binary silhouettes from segmented images processed by the first two segmentation networks. The NN generated pose estimates were inputs to a numerical model-image registration optimization routine. The results of completely autonomous measurements were compared with human-supervised measurements.

Results: The RMS differences between autonomous and human-supervised kinematics were less than 0.3mm or 0.4° for knee joint translations or rotations.Conclusions: With a homogenous training dataset from a single study, the ML solutions with a numerical optimizer provide excellent results for autonomous measurement of dynamic 3D knee replacement kinematics. Future work will generalize these capabilities to heterogeneous image sources containing different imaging systems and implant components. This pilot study demonstrates the potential to generate accurate joint kinematics measurements autonomously, suggesting a clinically practical, inexpensive method for quantifying joint mechanics.