
Medical device-related pressure injury remains a concern in harness-based body weight support systems because interface pressure alone does not characterize internal tissue loading. To design harness-based interfaces that balance load transfer and tissue safety, this study developed a mechanics-driven framework integrating CLO 3D pressure mapping, surface-driven finite element (FE) modeling, harness structure synthesis, and regional material optimization. Four representative harness structures were evaluated in models derived from 14 male post-stroke participants across five independent quasi-static gait postures, and favorable structural features were synthesized into a candidate configuration. For this candidate, 16 region-specific material variables were optimized using an optimal Latin hypercube design followed by adaptive simulated annealing. The resulting prototype was fabricated and evaluated against three clinical harnesses in an independent cohort of 16 male post-stroke participants. The optimal prototype reduced mean pressure and peak pressure across regions by 8.16-50.92% and 27.08-66.23%, respectively. FE analysis showed lower skin von Mises stress and principal shear strain than R1 and R2 across all regions.
Ultrasound (US) imaging is central to the assessment of musculoskeletal disorders, but robotic systems still rely on operators for initial joint localization and scan-region selection, limiting standardization. We present an AI-guided robotic US framework that combines two-step deep-learning alignment with optical refraction compensation for underwater joint scanning. The fully automated workflow performs underwater localization, per-joint US probe centering, and inter-joint path generation. Within 2 minutes, it captures cross-sectional and sagittal images of a whole hand, covering five metacarpophalangeal joints, the thumb interphalangeal joint, and four proximal interphalangeal joints. The alignment network estimates the optimal scan center with 1–2 mm accuracy despite initial errors up to 1 cm. In static-target validation, the optical model reduces refraction-induced localization error by 88%. Independent expert grading of 200 joint-orientation sweeps yielded a mean image-quality score of 3.33±1.08/5, and 77.5% were rated good or better.
Localization of miniature magnets inside the body could enable radiation-free guidance for endoscopy, catheter navigation and robotic interventions. However, existing systems often rely on relatively large magnets that generate strong signals at the sensor array. Here, we present a system-level study of AI-based pose estimation for a small magnet using Hall-effect sensor measurements. We compare physics-based dipole inversion, conventional machine-learning models and deep neural networks, and use neural architecture search to develop a task-specific model termed magNET. Evaluation on a large, robotically generated benchtop dataset using conventional 80/20 splits and structured hold-out experiments showed that neural network-based models outperformed the analytical and rule-based baselines. Among the learning-based approaches, magNET and a residual multilayer perceptron achieved the strongest and broadly comparable performance. magNET provided modest improvements in several conditions, including reductions in position error of approximately 10% and orientation error of up to approximately 30%, although these gains were not consistent across all structured hold-outs. Within the central sensor-board footprint and at elevations between 40 and 80 mm, pose-estimation errors remained below 2 mm and 2° under structured hold-out evaluation. Errors increased outside this region as signal quality declined. Model inference time was below 3 ms, although end-to-end real-time performance remains to be established. Overall, the results demonstrate the potential of Hall-effect sensing combined with neural networks for accurate localization of miniature magnets and future catheter-tracking applications.
Laryngeal cancer is a malignancy of the upper aerodigestive tract. Outpatient screening and transoral laser microsurgery for the treatment of laryngeal cancer rely on endoscopic imaging, where surgeons interpret complex visual data for diagnosis and margin assessment. Deep learning methods can support these tasks, but they typically assume high-quality input frames, a condition rarely satisfied in clinical practice. Automatically selecting informative frames may help reduce false-positive cancer detections and computational redundancy. However, training a model for anatomically relevant frame selection in a fully supervised fashion is challenging, as labeling is not part of the clinical workflow and is time-consuming. Therefore, we propose an active learning framework based on DINOv3 embeddings to identify informative frames while minimizing annotation effort through iterative selection. Our method achieves performance comparable to standard approaches (F1-score of 0.85), but requires less than 10% of the number of annotated frames they typically need (up to 1000). Furthermore, we introduce a personalization strategy to adapt the framework to videos acquired in different settings, and validate our method on 110 videos from 110 patients from two clinical centers. Such results pave the way for translating computer-assisted algorithms for laryngeal disease screening and margin assessment into surgical practice.
Patient-image registration is a core component of orthopedic navigation systems because it directly determines the overall precision of surgical navigation. Most existing methods adopt a coarse-to-fine paradigm, in which coarse registration provides a reliable initialization for subsequent fine alignment. However, conventional coarse registration methods require multiple bony landmarks, which are difficult to localize accurately owing to overlying soft tissue, resulting in substantial errors and necessitating repeated localization attempts. To address this limitation, we propose a novel coarse registration method based on the inherent rotation center of bone structures. By exploiting the rotation center, the translation component can be recovered, thereby reducing the registration to a pure rotation estimation problem. Furthermore, a rotation estimation strategy using interval stabbing is developed. The proposed method does not require multiple bony landmarks; instead, only a rotation center is needed to achieve coarse registration. Extensive experiments demonstrate that the proposed method provides more accurate initialization for fine registration. Specifically, it achieves registration errors of 2.18±0.81 mm and 1.96±0.72 mm on the acetabulum and femur phantoms, respectively. These results indicate that the proposed approach has strong potential for clinical applications.