Attaining venous access is a common requirement for clinical care worldwide, with a non-negligible portion of cases often being categorized as ‘difficult intravenous access’. Such complications result in far-reaching consequences affecting clinicians and patients alike. We propose a mixed-reality-based vein detection and visual guidance system that provides several key advantages, including a wider field of view, flexible operating distance, and hands-free, intuitive usage compared to existing solutions. A semi-supervised learning approach was used in model training to circumvent dataset availability limitations. Quantitative evaluation showed that the semi-supervised approach improved vein detection performance and temporal consistency. The system was also implemented and trialed in a clinical setting to assess real-world usability. Initial, preliminary assessment of HoloVein by medical professionals in a clinical setting showed improvements in detection quality using the semi-supervised approach over the baseline model. This result was deemed to be promising from a clinical perspective and could set the stage for more widespread mixed-reality venipuncture guidance tools in the future.

State-of-the-art image matching methods have shown strong generalization across natural image datasets, but their effectiveness in complex surgical environments remains underexplored. Surgical scenes introduce unique challenges, including homogeneous tissue textures, variable lighting, and frequent occlusions, which can degrade the reliability of keypoint correspondences essential for downstream vision tasks such as camera pose estimation and structure-from-motion. In this study, we systematically evaluate leading image matching methods within laparoscopic surgical settings, emphasizing performance under resource-constrained conditions. We present an optimized evaluation pipeline that incorporates robust estimators to enhance correspondence filtering and assess their impact on pose estimation accuracy. Our approach also examines the influence of fine-tuning individual pipeline components, particularly robust estimators, on overall system performance. Mean Reprojection error is refined by thresholding the nearest ground truth projections, enabling a more precise characterization of matching accuracy. Across five robust estimators, FM_8PTS consistently demonstrates superior resilience to outliers. Our results establish RoMa as the leading model for balancing pose estimation accuracy, reprojection performance, and computational efficiency, making it suitable for real-time surgical applications. By providing the first systematic benchmark and actionable insights for optimizing image matching pipelines in surgical domains, this work sets a new standard and paves the way for more reliable, efficient, and clinically applicable image-guided tools in minimally invasive surgery.

Ultrasound (US) imaging provides a safe and accessible solution to procedural guidance and diagnostic imaging. The effective usage of conventional 2D US for interventional guidance requires extensive experience to project the image plane onto the patient, and the interpretation of images in diagnostics suffers from high intra- and inter-user variability. 3D US reconstruction allows for more consistent diagnosis and interpretation, but existing solutions are limited in terms of equipment and applicability in real-time navigation. To address these issues, we propose HoloPOCUS - a mixed reality US system (MR-US) that overlays rich US information onto the user's vision in a point-of-care setting. HoloPOCUS extends existing MR-US methods beyond placing a US plane in the user's vision to include a 3D reconstruction and projection that can aid in procedural guidance using conventional probes. We validated a tracking pipeline that demonstrates higher accuracy compared to existing MR-US works. Furthermore, user studies conducted via a phantom task showed significant improvements in navigation duration when using our proposed methods.

Background: Deep learning models have the potential to enable fast and consistent interpretations of scoliosis radiographs. This study aims to assess the impact of deep learning assistance on the speed and accuracy of clinicians in measuring major coronal curves on scoliosis radiographs. Methods: We utilized a deep learning model (Context Axial Reverse Attention Network, or CaraNet) to assist in measuring Cobb’s angles on scoliosis radiographs in a simulated clinical setting. Four trainee radiologists with no prior experience and four trainee orthopedists with four to six months of prior experience analyzed the radiographs retrospectively, both with and without deep learning assistance, using a six-week washout period. We recorded the interpretation time and mean angle differences, with a consultant spine surgeon providing the reference standard. The dataset consisted of 640 radiographs from 640 scoliosis patients, aged 10–18 years; we divided the dataset into 75% for training, 16% for validation, and 9% for testing. Results: Deep learning assistance achieved non-statistically significant improvements in mean accuracy of 0.32 for trainee orthopedists (95% CI −1.4 to 0.8, p > 0.05) and 0.43 degrees (95% CI −1.6 to 0.8, p > 0.05) for trainee radiologists (non-inferior across all readers). Mean interpretation time decreased by 13.25 s for trainee radiologists, but increased by 3.85 s for trainee orthopedists (p = 0.005). Conclusions: Deep learning assistance for measuring Cobb’s angles was as accurate as unaided interpretation and slightly improved measurement accuracy. It increased the interpretation speeds of trainee radiologists but slightly slowed trainee orthopedists, suggesting that its effect on speed depended on prior experience.