ABSTRACT. Ultrasound (US) bone segmentation is an important component of US-guided orthopaedic procedures. While there are many published segmentation techniques, there is no direct way to compare their performance. We present a solution to this, by curating a multi-institutional set of US images and corresponding segmentations, and systematically evaluating six previously-published bone segmentation algorithms using consistent metric definitions. We find that learning-based segmentation methods outperform traditional algorithms that rely on hand-crafted image features, as measured by their Dice scores, RMS distance errors and segmentation success rates. However, there is no single best performing algorithm across the datasets, emphasizing the need for carefully evaluating techniques on large, heterogenous datasets. The datasets and evaluation framework described can be used to accelerate development of new segmentation algorithms.

ABSTRACT. Purpose Distal radius fractures (DRF) are common types of fractures with a high incident rate. DRF can be treated either by cast or surgery. To determine the clinical procedure and the operative management, standardized guidelines have become increasingly common. As operative indications are controversial, radiographic parameters (RPs) can provide objective support for effective decision making. Calculating the RPs manually from radiographs is time consuming and subject to observer variability and clinician experience. Our aim was to develop an automatic method for accurately and reliably computing 10 RPs associated with DRF in anteroposterior (AP) and lateral radiographs of a fractured hand with and without cast. Methods The inputs are the AP and lateral radiographs of the fractured hand with or without cast. The outputs are 10 RP values and composite images showing the landmark points and axes used in the RPs computation on the radiographs. Our method comprises three main steps: 1) segmentation of the radius and the ulna with a deep learning radiograph pixel classifier; 2) landmark points and axis extraction from the segmentations using geometric model-based methods; 3) RPs computation from the landmarks and generation of composite images. Our study tested the accuracy of step 2. The dataset consists of 20 pairs of AP and lateral radiographs. Ground truth radius and ulna segmentations were manually performed by an expert clinician co-author. Ground truth landmarks were manually located and annotated by the two expert clinician co-authors. The computed RP was considered accurate (in range) when its value was inside the inter and intra observer variability range of the manual annotation. The overall accuracy of the AP and lateral measurements was obtained by averaging the accuracy of each RP. Results The accuracy of the computed AP RPs is 92.7%. The Radial Length and Radial Shift are within the observer variability range; for the Radial Angle, Ulnar Variance and Step all cases are within range except for one outlier; the Gap has two outlier cases. The accuracy of the computed lateral RPs is 100%: all four Palmer Tilt, Dorsal Shift, Gap, and Step are within the clinician observer variability.

Conclusion Automatic computation of distal radius fractures RPs from AP and lateral radiographs of hands with and without cast can be performed accurately. Precise and consistent measurement of RPs may improve the clinical decision making process.

ABSTRACT. Progress in machine learning and artificial intelligence (AI) opens the way to the development of smart clinical-assistance systems and decision-support tools for the operating room (OR). Yet, before deploying these algorithms in the OR, assessment of their performances in real clinical conditions is necessary. Gathering intraoperative data for training and testing is hard, and robustness to the challenging conditions of the OR is not always demonstrated. In this paper we introduce a unique multi-patient dataset of images captured during Total Knee Arthroplasty (TKA) surgery. We use this dataset to compare five deep learning-based image segmentation approaches and provide quantitative and qualitative results. We hope that this work will help bringing light on the performances of AI in a real surgical environment.

ABSTRACT. Fully-automatic and reliable segmentation of bone surface in volumetric ultrasound images could enable the use of this imaging technique for a variety of tasks, including diagnosis of hip dysplasia, ACL injuries in the knee as well as patient-specific instrumentation and implants in total hip or knee arthroplasty. Interpretation of volumetric data is a hard task, even for humans. In this study, we investigate the benefit of using the spatial information of a third dimension on the task of segmentation of the distal femoral bone. A data set of 52 volumetric image with 12771 image slices is split into a training and test set. We employ 2D and 3D variants of the nnUNet architecture and compare the accuracy in terms of dice coefficient and performance in terms of inference time. Note that processing of 2D data allows for a bigger model due to less memory consumption. Both architectures achieve a Dice of about 82% while the 2D variant shows less false positive segmentation and achieves a surface distance error of 0.44mm, in contrast to 0.81mm for the 3D variant. At the same time, the former infers three times faster at about 10 seconds per volume image. Apparently, model size has a bigger positive effect than the additional spatial information. Thus, we recommend considering 2D segmentation architectures even for volumetric segmentation tasks.

ABSTRACT. The current standard for Computer-Assisted Orthopaedic Surgery is to use large optical trackers with a wide field-of-view, placing them outside the sterile field. The widespread availability of highly accurate depth sensing technology, however, allows to envision novel solutions for patient registration and tracking, which were previously unavailable. Indeed, the small size and light weight of many commercial RGBD sensors allows them to be worn by the surgeon or embedded into surgical tools. This work demonstrates a setup in which a compact RGBD sensor is mounted on a robotic tool for assisted bone preparation. Such a setup is appealing from a technical point of view because it would allow to implement robot assistance without needing to track any instrumentation. We coupled this setup with a deep-learning based algorithm for automated segmentation and tracking of patient anatomy, and tested the accuracy of patient registration on a synthetic knee anatomical model, using a marker-based optical tracker as gold standard reference.

ABSTRACT. In order to follow up volume changes of thigh muscles in either disorders or muscle therapy treatments, several segmentation methods have been introduced. Since the accuracy of such methods is crucial to assess muscle reconstruction, we introduced a semi-automatic tool to segment thigh muscles with the required accuracy. The tool segments each muscle in three steps. First, a few slices are manually annotated for each muscle. Then, all of these annotated contours are automatically connected using spline interpolation curves in both transversal and longitudinal directions. Finally, using morphological and image processing techniques, each 3D muscle is reconstructed in order to further analyze their volumes. Five healthy subjects were invited to scan their thighs (10 tight). Seven muscles from each thigh were selected: the Rectus Femoris, Vastus Lateralis, Vastus Intermedius, Vastus Medialis, Biceps Short Head, Semitendinosus and Semimembranosus (70 muscles in total). We evaluated our method using fully manual ground truth as well as well-known available tool in 3D Slicer. The results showed that the accuracy and executing time of our tool are compatible with the clinical requirements.

ABSTRACT. The automated and robust segmentation of bone surfaces in ultrasound (US) images can open up new fields of application for US imaging in computer-assisted orthopedic surgery, e.g. for the patient-specific planning process in computer-assisted knee replacement. For the automated, deep learning-based segmentation of medical images, CNN-based methods have been the state of the art over the last years, while recently Transformer-based methods are on the rise in computer vision. To compare these methods with respect to US image segmentation, in this paper the recent Transformer-based Swin-UNet is exemplarily benchmarked against the commonly used CNN-based nnUNet on the application of in-vivo 2D US knee segmentation. Trained and tested on our own dataset with 8166 annotated images (split in 7155 and 1011 images respectively), both the nnUNet and the pre-trained Swin-UNet show a Dice coefficient of 0.78 during testing. For distances between skeletonized labels and predictions, a symmetric Hausdorff distance of 44.69 pixels and a symmetric surface distance of 5.77 pixels is found for nnUNet as compared to 42.78 pixels and 5.68 pixels respectively for the Swin-UNet. Based on qualitative assessment, the Transformer-based Swin-UNet appears to benefit from its capability of learning global relationships as compared to the CNN-based nnUNet, while the latter shows more consistent and smooth predictions on a local level, presumably due to the character of convolution operation. Besides, the Swin-UNet requires generalized pre-training to be competitive. Since both architectures are evenly suited for the task at hand, for our future work, hybrid architectures combining the characteristic advantages of Transformer-based and CNN-based methods seem promising for US image segmentation.

ABSTRACT. Reference axis based on Friedman’s approach is widely recognized as an anatomic landmark allowing to measure and compare implant parameters within preoperative planning software for total shoulder arthroplasty. Equinoxe Planning Application (ExactechInc.) offers 3D measurements techniques for glenoid version and inclination requiring meticulous placement of trigonum and glenoid center. We propose an automatic determination of this reference axis, based on deep learning that shown a median error lower than 1°.

ABSTRACT. Two-year minimum clinical outcomes were collected on anatomic and reverse total shoulder arthroplasty patients enrolled in a single implant global registry that were performed using an intraoperative computer navigated surgery system. Age, gender, and follow-up matched cohorts were created from the same registry for comparison purposes for both anatomic and reverse total shoulder arthroplasty. The navigated cohorts exhibited as good or better clinical outcomes compared to the non-navigated cohorts as well as reductions in postoperative complications, revision rates, and adverse events.

ABSTRACT. Rotator cuff tears (RCT) are one of the most common sources of shoulder pain. Many factors can be considered to choose the right surgical treatment procedure. Of the most important factors are the tear retraction and tear width, assessed manually on preoperative MRI. A novel approach to automatically quantify a rotator cuff tear, based on the segmentation of the tear from MRI images, was developed and validated. For segmentation, a neural network was trained and methods for the automatic calculation of the tear width and retraction from the segmented tear volume were developed. The accuracy of the automatic segmentation and the automated tear analysis were evaluated relative to manual consensus segmentations by two clinical experts. Variance in the manual segmentations was assessed in an interrater variability study of two clinical experts. The accuracy of the tear retraction calculation based on the developed automatic tear segmentation was 5.3 mm ± 5.0 mm in comparison to the interrater variability of tear retraction calculation based on manual segmentations of 3.6 mm ± 2.9 mm. These results show that an automatic quantification of a rotator cuff tear is possible. The large interrater variability of manual segmentation-based measurements highlights the difficulty of the tear segmentations task in general.

ABSTRACT. Preoperative anatomic measurements in total shoulder arthroplasty (TSA) influence a surgeon’s decision-making process in deciding treatment options for a given patient. Glenoid retroversion is one of the most significant measurements and can be highly subject to intra- and inter-observer variability in measurement technique. This study compares surgeon measured retroversion values to semi-automated software measured retroversion values on the same 1862 computed tomography scans, showing consistent measurements with an average absolute mean error between the two techniques of 3.1 ± 3.6°.

ABSTRACT. The primary objective of this study was to obtain a reliable method of automatic segmentation of shoulder muscles. The secondary objective of this study was to define a new computed tomography (CT)-based quantitative 3-dimensional (3D) measure of muscle loss (3DML) based on the rationale of the 2-dimensional (2D) qualitative Goutallier score. 102 CT scans were manually segmented and an algorithm of automated segmentation of the muscles was created. The volume of muscle fibers without intramuscular fat was then calculated for each rotator cuff muscle and normalized to the patient's scapular volume to account for the effect of body size (NVfibers). 3D muscle mass (3DMM) was calculated by dividing the NVfibers value of a given muscle by the mean expected volume in healthy shoulders. 3D muscle loss (3DML) was defined as 1 - (3DMM). Automated segmentation of the muscles was possible with a mean Dice of 0.904 ± 0.01 for the deltoid, 0.887 ± 0.014 for the infraspinatus (ISP), 0.892 ± 0.008 for the subscapularis (SSC), 0.885 for the supraspinatus (SSP) and 0.796 ± 0.006 for the teres minor (TM). The mean values of 3DFI and 3DML were 0.9% and 5.3% for Goutallier 0, 2.9% and 25.6% for Goutallier 1, 11.4% and 49.5% for Goutallier 2, 20.7% and 59.7% for Goutallier 3, and 29.3% and 70.2% for Goutallier 4, respectively. 3DML measurements obtained automatically incorporate both atrophy and fatty infiltration, thus they could become a very reliable index for assessing shoulder muscle function which could help in the decision process in shoulder surgery.

ABSTRACT. Fat fraction of the rotator cuff muscles has been shown to be a predictor of rotator cuff repair failure. In clinical diagnosis, fat fraction of the affected muscle is typically assessed visually on the oblique 2D Y-view and categorized according to the Goutallier scale on T1 weighted MRI. To enable a quantitative fat fraction measure of the rotator cuff muscles, an automated analysis of the whole muscle and Y-view slice was developed utilizing 2-point Dixon MRI. 3D nn-Unet were trained on water only 2-point Dixon data and corresponding annotations for the automatic segmentation of the supraspinatus, humerus and scapula and the detection of 3 anatomical landmarks for the automatic reconstruction of the Y-view slice. The supraspinatus was segmented with a Dice coefficient of 90% (N=24) and automatic fat fraction measurements with a difference from manual measurements of 1.5 % for whole muscle and 0.6% for Y-view evaluation (N=21) were observed. The presented automatic analysis demonstrates the feasibility of a 3D quantification of fat fraction of the rotator cuff muscles for the investigation of more accurate predictors of rotator cuff repair outcome.