Geometric deep learning is a relatively new field that combines the principles of deep learning with techniques from geometry and topology to analyze data with complex structures, such as graphs and manifolds. In orthopedic research, geometric deep learning has been applied to a variety of tasks, including the analysis of imaging data to detect and classify abnormalities, the prediction of patient outcomes following surgical interventions, and the identification of risk factors for degenerative joint disease. This review aims to summarize the current state of the field and highlight the key findings and applications of geometric deep learning in orthopedic research. The review also discusses the potential benefits and limitations of these approaches and identifies areas for future research. Overall, the use of geometric deep learning in orthopedic research has the potential to greatly advance our understanding of the musculoskeletal system and improve patient care

Distal radius fractures (DRFs) are one of the most common types of fracture and one which is often treated surgically. Standard X-rays are obtained for DRFs, and in most cases that have an intra-articular component, a routine CT is also performed. However, it is estimated that CT is only required in 20% of cases and therefore routine CT's results in the overutilisation of resources burdening radiology and emergency departments. In this study, we explore the feasibility of using deep learning to differentiate intra- and extra-articular DRFs automatically and help streamline which fractures require a CT. Retrospectively x-ray images were retrieved from 615 DRF patients who were treated with an ORIF at the Royal Brisbane and Women's Hospital. The images were classified into AO Type A, B or C fractures by three training registrars supervised by a consultant. Deep learning was utilised in a two-stage process: 1) localise and focus the region of interest around the wrist using the YOLOv5 object detection network and 2) classify the fracture using a EfficientNet-B3 network to differentiate intra- and extra-articular fractures. The distal radius region of interest (ROI) detection stage using the ensemble model of YOLO networks detected all ROIs on the test set with no false positives. The average intersection over union between the YOLO detections and the ROI ground truth was Error! Digit expected.. The DRF classification stage using the EfficientNet-B3 ensemble achieved an area under the receiver operating characteristic curve of 0.82 for differentiating intra-articular fractures. The proposed DRF classification framework using ensemble models of YOLO and EfficientNet achieved satisfactory performance in intra- and extra-articular fracture classification. This work demonstrates the potential in automatic fracture characterization using deep learning and can serve to streamline decision making for axial imaging helping to reduce unnecessary CT scans

Abstract. OBJECTIVES. Application of deep learning approaches to marker trajectories and ground reaction forces (mocap data), is often hampered by small datasets. Enlarging dataset size is possible using some simple numerical approaches, although these may not be suited to preserving the physiological relevance of mocap data. We propose augmenting mocap data using a deep learning architecture called “generative adversarial networks” (GANs). We demonstrate appropriate use of GANs can capture variations of walking patterns due to subject- and task-specific conditions (mass, leg length, age, gender and walking speed), which significantly affect walking kinematics and kinetics, resulting in augmented datasets amenable to deep learning analysis approaches. METHODS. A publicly available (. https://www.nature.com/articles/s41597-019-0124-4. ) gait dataset (733 trials, 21 women and 25 men, 37.2 ± 13.0 years, 1.74 ± 0.09 m, 72.0 ± 11.4 kg, walking speeds ranging from 0.18 m/s to 2.04 m/s) was used as the experimental dataset. The GAN comprised three neural networks: an encoder, a decoder, and a discriminator. The encoder compressed experimental data into a fixed-length vector, while the decoder transformed the encoder's output vector and a condition vector (containing information about the subject and trial) into mocap data. The discriminator distinguished between the encoded experimental data from randomly sampled vectors of the same size. By training these networks jointly using the experimental dataset, the generator (decoder) could generate synthetic data respecting specified conditions from randomly sampled vectors. Synthetic mocap data and lower limb joint angles were generated and compared to the experimental data, by identifying the statistically significant differences across the gait cycle for a randomly selected subset of the experimental data from 5 female subjects (73 trials, aged 26–40, weighing 57–74 kg, with leg lengths between 868–931 mm, and walking speeds ranging from 0.81–1.68 m/s). By conducting these comparisons for this subset, we aimed to assess the synthetic data generated using multiple conditions. RESULTS. We visually inspected the synthetic trials to ensure that they appeared realistic. The statistical comparison revealed that, on average, only 2.5% of the gait cycle showed significantly differences in the joint angles of the two data groups. Additionally, the synthetic ground reaction forces deviated from the experimental data distribution for an average of 2.9% of the gait cycle. CONCLUSIONS. We introduced a novel approach for generating synthetic mocap data of human walking based on the conditions that influence walking patterns. The synthetic data closely followed the trends observed in the experimental data, also in the literature, suggesting that our approach can augment mocap datasets considering multiple conditions, an approach unfeasible in previous work. Creation of large, augmented datasets allows the application of other deep learning approaches, with the potential to generate realistic mocap data from limited and non-lab-based data. Our method could also enhance data sharing since synthetic data does not raise ethical concerns. You can generate and download virtual gait data using our GAN approach from . https://thisgaitdoesnotexist.streamlit.app/. . Declaration of Interest. (b) declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the research reported:I declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the research project

Introduction. Three-dimensional (3D) morphological understanding of the hip joint, specifically the joint space and surrounding anatomy, including the proximal femur and the pelvis bone, is crucial for a range of orthopedic diagnoses and surgical planning. While deep learning algorithms can provide higher accuracy for segmenting bony structures, delineating hip joint space formed by cartilage layers is often left for subjective manual evaluation. This study compared the performance of two state-of-the-art 3D deep learning architectures (3D UNET and 3D UNETR) for automated segmentation of proximal femur bone, pelvis bone, and hip joint space with single and multi-class label segmentation strategies. Method. A dataset of 56 3D CT images covering the hip joint was used for the study. Two bones and hip joint space were manually segmented for training and evaluation. Deep learning models were trained and evaluated for a single-class approach for each label (proximal femur, pelvis, and the joint space) separately, and for a multi-class approach to segment all three labels simultaneously. A consistent training configuration of hyperparameters was used across all models by implementing the AdamW optimizer and Dice Loss as the primary loss function. Dice score, Root Mean Squared Error, and Mean Absolute Error were utilized as evaluation metrics. Results. Both the models performed at excellent levels for single-label segmentations in bones (dice > 0.95), but single-label joint space performance remained considerably lower (dice < 0.87). Multi-class segmentations remained at lower performance (dice < 0.88) for both models. Combining bone and joint space labels may have introduced a class imbalance problem in multi-class models, leading to lower performance. Conclusion. It is not clear if 3D UNETR provides better performance as the selection of hyperparameters was the same across the models and was not optimized. Further evaluations will be needed with baseline UNET and nnUNET modeling architectures

Evaluation of patient specific spinopelvic mobility requires the detection of bony landmarks in lateral functional radiographs. Current manual landmarking methods are inefficient, and subjective. This study proposes a deep learning model to automate landmark detection and derivation of spinopelvic measurements (SPM). A deep learning model was developed using an international multicenter imaging database of 26,109 landmarked preoperative, and postoperative, lateral functional radiographs (HREC: Bellberry: 2020-08-764-A-2). Three functional positions were analysed: 1) standing, 2) contralateral step-up and 3) flexed seated. Landmarks were manually captured and independently verified by qualified engineers during pre-operative planning with additional assistance of 3D computed tomography derived landmarks. Pelvic tilt (PT), sacral slope (SS), and lumbar lordotic angle (LLA) were derived from the predicted landmark coordinates. Interobserver variability was explored in a pilot study, consisting of 9 qualified engineers, annotating three functional images, while blinded to additional 3D information. The dataset was subdivided into 70:20:10 for training, validation, and testing. The model produced a mean absolute error (MAE), for PT, SS, and LLA of 1.7°±3.1°, 3.4°±3.8°, 4.9°±4.5°, respectively. PT MAE values were dependent on functional position: standing 1.2°±1.3°, step 1.7°±4.0°, and seated 2.4°±3.3°, p< 0.001. The mean model prediction time was 0.7 seconds per image. The interobserver 95% confidence interval (CI) for engineer measured PT, SS and LLA (1.9°, 1.9°, 3.1°, respectively) was comparable to the MAE values generated by the model. The model MAE reported comparable performance to the gold standard when blinded to additional 3D information. LLA prediction produced the lowest SPM accuracy potentially due to error propagation from the SS and L1 landmarks. Reduced PT accuracy in step and seated functional positions may be attributed to an increased occlusion of the pubic-symphysis landmark. Our model shows excellent performance when compared against the current gold standard manual annotation process

Introduction. Inaccurate identification of implants on X-rays may lead to prolonged surgical duration as well as increased complexity and costs during implant removal. Deep learning models may help to address this problem, although they typically require large datasets to effectively train models in detecting and classifying objects, e.g. implants. This can limit applicability for instances when only smaller datasets are available. Transfer learning can be used to overcome this limitation by leveraging large, publicly available datasets to pre-train detection and classification models. The aim of this study was to assess the effectiveness of deep learning models in implant localisation and classification on a lower limb X-ray dataset. Method. Firstly, detection models were evaluated on their ability to localise four categories of implants, e.g. plates, screws, pins, and intramedullary nails. Detection models (Faster R-CNN, YOLOv5, EfficientDet) were pre-trained on the large, freely available COCO dataset (330000 images). Secondly, classification models (DenseNet121, Inception V3, ResNet18, ResNet101) were evaluated on their ability to classify five types of intramedullary nails. Localisation and classification accuracy were evaluated on a smaller image dataset (204 images). Result. The YOLOv5s model showed the best capacity to detect and distinguish between different types of implants (accuracy: plate=82.1%, screw=72.3%, intramedullary nail=86.9%, pin=79.9%). Screw implants were the most difficult implant to detect, likely due to overlapping screw implants visible in the image dataset. The DenseNet121 classification model showed the best performance in classifying different types of intramedullary nails (accuracy=73.7%). Therefore, a deep learning model pipeline with the YOLOv5s and DenseNet121 was proposed for the most optimal performance of automating implants localisation and classification for a relatively small dataset. Conclusion. These findings support the potential of deep learning techniques in enhancing implant detection accuracy. With further development, AI-based implant identification may benefit patients, surgeons and hospitals through improved surgical planning and efficient use of theatre time

Abstract. Objectives. Knee alignment affects both the development and surgical treatment of knee osteoarthritis. Automating femorotibial angle (FTA) and hip-knee-ankle angle (HKA) measurement from radiographs could improve reliability and save time. Further, if the gold-standard HKA from full-limb radiographs could be accurately predicted from knee-only radiographs then the need for more expensive equipment and radiation exposure could be reduced. The aim of this research is to assess if deep learning methods can predict FTA and HKA angle from posteroanterior (PA) knee radiographs. Methods. Convolutional neural networks with densely connected final layers were trained to analyse PA knee radiographs from the Osteoarthritis Initiative (OAI) database with corresponding angle measurements. The FTA dataset with 6149 radiographs and HKA dataset with 2351 radiographs were split into training, validation and test datasets in a 70:15:15 ratio. Separate models were learnt for the prediction of FTA and HKA, which were trained using mean squared error as a loss function. Heat maps were used to identify the anatomical features within each image that most contributed to the predicted angles. Results. FTA could be predicted with errors less than 3° for 99.8% of images, and less than 1° for 89.5%. HKA prediction was less accurate than FTA but still high: 95.7% within 3°, and 68.0 % within 1°. Heat maps for both models were generally concentrated on the knee anatomy and could prove a valuable tool for assessing prediction reliability in clinical application. Conclusions. Deep learning techniques could enable fast, reliable and accurate predictions of both FTA and HKA from plain knee radiographs. This could lead to cost savings for healthcare providers and reduced radiation exposure for patients

Introduction. Automated identification of arthroplasty implants could aid in pre-operative planning and is a task which could be facilitated through artificial intelligence (AI) and deep learning. The purpose of this study was to develop and test the performance of a deep learning system (DLS) for automated identification and classification of knee arthroplasty (KA) on radiographs. Methods. We collected 237 AP knee radiographs with equal proportions of native knees, total KA (TKA), and unicompartmental KA (UKA), as well as 274 radiographs with equal proportions of Smith & Nephew Journey and Zimmer NexGen TKAs. Data augmentation was used to increase the number of images available for DLS development. These images were used to train, validate, and test deep convolutional neural networks (DCNN) to 1) detect the presence of TKA; 2) differentiate between TKA and UKA; and 3) differentiate between the 2 TKA models. Receiver operating characteristic (ROC) curves were generated with area under the curve (AUC) calculated to assess test performance. Results. The DCNNs trained to detect KA and to distinguish between TKA and UKA both achieved AUC of 1. In both cases, heatmap analysis demonstrated appropriate emphasis of the KA components in decision-making. The DCNN trained to distinguish between the 2 TKA models also achieved AUC of 1. Heatmap analysis of this DCNN showed emphasis of specific unique features of the TKA model designs for decision making, such as the anterior flange shape of the Zimmer NexGen TKA (Figure 1) and the tibial baseplate/stem shape of the Smith & Nephew Journey TKA (Figure 2). Conclusion. DCNNs can accurately identify presence of TKA and distinguish between specific designs. The proof-of-concept of these DCNNs may set the foundation for DCNNs to identify other prosthesis models and prosthesis-related complications. For any figures or tables, please contact the authors directly

Introduction. Gait laboratory measurement of whole-body kinematics and ground reaction forces during a wide range of activities is frequently performed in joint replacement patient diagnosis, monitoring, and rehabilitation programs. These data are commonly processed in musculoskeletal modeling platforms such as OpenSim and Anybody to estimate muscle and joint reaction forces during activity. However, the processing required to obtain musculoskeletal estimates can be time consuming, requires significant expertise, and thus seriously limits the patient populations studied. Accordingly, the purpose of this study was to evaluate the potential of deep learning methods for estimating muscle and joint reaction forces over time given kinematic data, height, weight, and ground reaction forces for total knee replacement (TKR) patients performing activities of daily living (ADLs). Methods. 70 TKR patients were fitted with 32 reflective markers used to define anatomical landmarks for 3D motion capture. Patients were instructed to perform a range of tasks including gait, step-down and sit-to-stand. Gait was performed at a self-selected pace, step down from an 8” step height, and sit-to-stand using a chair height of 17”. Tasks were performed over a force platform while force data was collected at 2000 Hz and a 14 camera motion capture system collected at 100 Hz. The resulting data was processed in OpenSim to estimate joint reaction and muscle forces in the hip and knee using static optimization. The full set of data consisted of 135 instances from 70 patients with 63 sit-to-stands, 15 right-sided step downs, 14 left-sided step downs, and 43 gait sequences. Two classes of neural networks (NNs), a recurrent neural network (RNN) and temporal convolutional neural network (TCN), were trained to predict activity classification from joint angle, ground reaction force, and anthropometrics. The NNs were trained to predict muscle and joint reaction forces over time from the same input metrics. The 135 instances were split into 100 instances for training, 15 for validation, and 20 for testing. Results. The RNN and TCN yielded classification accuracies of 90% and 100% on the test set. Correlation coefficients between ground truth and predictions from the test set ranged from 0.81–0.95 for the RNN, depending on the activity. Predictions from both NNs were qualitatively assessed. Both NNs were able to effectively learn relationships between the input and output variables. Discussion. The objective of the study was to develop and evaluate deep learning methods for predicting patient mechanics from standard gait lab data. The resulting models classified activities with excellent performance, and showed promise for predicting exact values for loading metrics for a range of different activities. These results indicate potential for real-time prediction of musculoskeletal metrics with application in patient diagnostics and rehabilitation. For any figures or tables, please contact authors directly

Quantitative ultrasound (QUS) is a promising tool to estimate bone structure characteristics and predict fragile fracture. The aim of this pilot cross-sectional study was to evaluate the performance of a multi-channel residual network (MResNet) based on ultrasonic radiofrequency (RF) signal to discriminate fragile fractures retrospectively in postmenopausal women. Methods. RF signal and speed of sound (SOS) were obtained using an axial transmission QUS at one‐third distal radius for 246 postmenopausal women. Based on the involved RF signal, we conducted a MResNet, which combines multi-channel training with original ResNet, to classify the high risk of fragility fractures patients from all subjects. The bone mineral density (BMD) at lumber, hip and femoral neck acquired with DXA was recorded on the same day. The fracture history of all subjects in adulthood were collected. To assess the ability of the different methods in the discrimination of fragile fracture, the odds ratios (OR) calculated using binomial logistic regression analysis and the area under the receiver operator characteristic curves (AUC) were analyzed. Results. Among the 246 postmenopausal women, 170 belonged to the non-fracture group, 50 to the vertebral group, and 26 to the non-vertebral fracture group. MResNet was discriminant for all fragile fractures (OR = 2.64; AUC = 0.74), for Vertebral fracture (OR = 3.02; AUC = 0.77), for non-vertebral fracture (OR = 2.01; AUC = 0.69). MResNet showed comparable performance to that of BMD of hip and lumbar with all types of fractures, and significantly better performance than SOS all types of fractures. Conclusions. the MResNet model based on the ultrasonic RF signal can significantly improve the ability of QUS device to recognize previous fragile fractures. Moreover, the performance of the proposed model modified by age, weight, and height is further optimized. These results open perspectives to evaluate the risk of fragile fracture applying a deep learning model to analyze ultrasonic RF signal

Introduction. The ability to walk over various surfaces such as cobblestones, slopes or stairs is a very patient centric and clinically meaningful mobility outcome. Current wearable sensors only measure step counts or walking speed regardless of such context relevant for assessing gait function. This study aims to improve deep learning (DL) models to classify surfaces of walking by altering and comparing model features and sensor configurations. Method. Using a public dataset, signals from 6 IMUs (Movella DOT) worn on various body locations (trunk, wrist, right/left thigh, right/left shank) of 30 subjects walking on 9 surfaces were analyzed (flat ground, ramps (up/down), stairs (up/down), cobblestones (irregular), grass (soft), banked (left/right)). Two variations of a CNN Bi-directional LSTM model, with different Batch Normalization layer placement (beginning vs end) as well as data reduction to individual sensors (versus combined) were explored and model performance compared in-between and with previous models using F1 scores. Result. The Bi-LSTM architecture improved performance over previous models, especially for subject-wise data splitting and when combining the 6 sensor locations (e.g. F1=0.94 versus 0.77). Placement of the Batch Normalization layer at the beginning, prior to the convolutional layer, enhanced model understanding of participant gait variations across surfaces. Single sensor performance was best on the right shank (F1=0.88). Conclusion. Walking surface detection using wearable IMUs and DL models shows promise for clinically relevant real-world applications, achieving high F1 levels (>0.9) even for subject-wise data splitting enhancing the model applicability in real-world scenarios. Normalization techniques, such as Batch Normalization, seem crucial for optimizing model performance across diverse participant data. Also single-sensor set-ups can give acceptable performance, in particular for specific surface types of potentially high clinical relevance (e.g. stairs, ramps), offering practical and cost-effective solutions with high usability. Future research will focus on collecting ground-truth labeled data to investigate system performance in real-world settings

Background

Dislocation is a common complication following total hip arthroplasty (THA), and accounts for a high percentage of subsequent revisions. The purpose of this study was to develop a convolutional neural network (CNN) model to identify patients at high risk for dislocation based on postoperative anteroposterior (AP) pelvis radiographs.

Aims

Total hip arthroplasty (THA) and total knee arthroplasty (TKA) are common orthopaedic procedures requiring postoperative radiographs to confirm implant positioning and identify complications. Artificial intelligence (AI)-based image analysis has the potential to automate this postoperative surveillance. The aim of this study was to prepare a scoping review to investigate how AI is being used in the analysis of radiographs following THA and TKA, and how accurate these tools are.

Aims

The number of convolutional neural networks (CNN) available for fracture detection and classification is rapidly increasing. External validation of a CNN on a temporally separate (separated by time) or geographically separate (separated by location) dataset is crucial to assess generalizability of the CNN before application to clinical practice in other institutions. We aimed to answer the following questions: are current CNNs for fracture recognition externally valid?; which methods are applied for external validation (EV)?; and, what are reported performances of the EV sets compared to the internal validation (IV) sets of these CNNs?

Aims

Hip dysplasia (HD) leads to premature osteoarthritis. Timely detection and correction of HD has been shown to improve pain, functional status, and hip longevity. Several time-consuming radiological measurements are currently used to confirm HD. An artificial intelligence (AI) software named HIPPO automatically locates anatomical landmarks on anteroposterior pelvis radiographs and performs the needed measurements. The primary aim of this study was to assess the reliability of this tool as compared to multi-reader evaluation in clinically proven cases of adult HD. The secondary aims were to assess the time savings achieved and evaluate inter-reader assessment.

Aims. The diagnosis of developmental dysplasia of the hip (DDH) is challenging owing to extensive variation in paediatric pelvic anatomy. Artificial intelligence (AI) may represent an effective diagnostic tool for DDH. Here, we aimed to develop an anteroposterior pelvic radiograph deep learning system for diagnosing DDH in children and analyze the feasibility of its application. Methods. In total, 10,219 anteroposterior pelvic radiographs were retrospectively collected from April 2014 to December 2018. Clinicians labelled each radiograph using a uniform standard method. Radiographs were grouped according to age and into ‘dislocation’ (dislocation and subluxation) and ‘non-dislocation’ (normal cases and those with dysplasia of the acetabulum) groups based on clinical diagnosis. The deep learning system was trained and optimized using 9,081 radiographs; 1,138 test radiographs were then used to compare the diagnoses made by deep learning system and clinicians. The accuracy of the deep learning system was determined using a receiver operating characteristic curve, and the consistency of acetabular index measurements was evaluated using Bland-Altman plots. Results. In all, 1,138 patients (242 males; 896 females; mean age 1.5 years (SD 1.79; 0 to 10) were included in this study. The area under the receiver operating characteristic curve, sensitivity, and specificity of the deep learning system for diagnosing hip dislocation were 0.975, 276/289 (95.5%), and 1,978/1,987 (99.5%), respectively. Compared with clinical diagnoses, the Bland-Altman 95% limits of agreement for acetabular index, as determined by the deep learning system from the radiographs of non-dislocated and dislocated hips, were -3.27° - 2.94° and -7.36° - 5.36°, respectively (p < 0.001). Conclusion. The deep learning system was highly consistent, more convenient, and more effective for diagnosing DDH compared with clinician-led diagnoses. Deep learning systems should be considered for analysis of anteroposterior pelvic radiographs when diagnosing DDH. The deep learning system will improve the current artificially complicated screening referral process. Cite this article: Bone Joint J 2020;102-B(11):1574–1581

The October 2023 Children’s orthopaedics Roundup. 360. looks at: Outcomes of open reduction in children with developmental hip dislocation: a multicentre experience over a decade; A torn discoid lateral meniscus impacts lower-limb alignment regardless of age; Who benefits from allowing the physis to grow in slipped capital femoral epiphysis?; Consensus guidelines on the management of musculoskeletal infection affecting children in the UK; Diagnosis of developmental dysplasia of the hip by ultrasound imaging using deep learning; Outcomes at a mean of 13 years after proximal humeral fracture during adolescence; Clubfeet treated according to Ponseti at four years; Controlled ankle movement boot provides improved outcomes with lower complications than short leg walking cast

The October 2023 Spine Roundup. 360. looks at: Cutting through surgical smoke: the science of cleaner air in spinal operations; Unlocking success: key factors in thoracic spine decompression and fusion for ossification of the posterior longitudinal ligament; Deep learning algorithm for identifying cervical cord compression due to degenerative canal stenosis on radiography; Surgeon experience influences robotics learning curve for minimally invasive lumbar fusion; Decision-making algorithm for the surgical treatment of degenerative lumbar spondylolisthesis of L4/L5; Response to preoperative steroid injections predicts surgical outcomes in patients undergoing fusion for isthmic spondylolisthesis

Patients with advanced cancer can develop bone metastases in the femur which are often painful and increase the risk of pathological fracture. Accurate segmentation of bone metastases is, amongst others, important to improve patient-specific computer models which calculate fracture risk, and for radiotherapy planning to determine exact radiation fields. Deep learning algorithms have shown to be promising to improve segmentation accuracy for metastatic lesions, but require reliable segmentations as training input. The aim of this study was to investigate the inter- and intra-operator reliability of manual segmentation of femoral metastatic lesions and to define a set of lesions which can serve as a training dataset for deep learning algorithms. F. CT-scans of 60 advanced cancer patients with a femur affected with bone metastases (20 osteolytic, 20 osteoblastic and 20 mixed) were used in this study. Two operators were trained by an experienced radiologist and then segmented the metastatic lesions in all femurs twice with a four-week time interval. 3D and 2D Dice coefficients (DCs) were calculated to quantify the inter- and intra-operator reliability of the segmentations. We defined a DC>0.7 as good reliability, in line with a statistical image segmentation study. Mean first and second inter-operator 3D-DCs were 0.54 (±0.28) and 0.50 (±0.32), respectively. Mean intra-operator I and II 3D-DCs were 0.56 (±0.28) and 0.71 (±0.23), respectively. Larger lesions (>60 cm. 3. ) scored higher DCs in comparison with smaller lesions. This study reveals that manual segmentation of metastatic lesions is challenging and that the current manual segmentation approach resulted in dissatisfying outcomes, particularly for lesions with small volumes. However, segmentation of larger lesions resulted in a good inter- and intra-operator reliability. In addition, we were able to select 521 slices with good segmentation reliability that can be used to create a training dataset for deep learning algorithms. By using deep learning algorithms, we aim for more accurate automated lesion segmentations which might be used in computer modelling and radiotherapy planning