International Journal of Computer Assisted Radiology and Surgery最新文献

Purpose: Surgical video review is essential for minimally invasive surgical training, but manual annotation of surgical steps is time-consuming and limits scalability. We propose a weakly supervised pre-training framework that leverages unannotated or heterogeneously labeled surgical videos to improve automated surgical step recognition.

Methods: We evaluate three types of weak labels derived from unannotated datasets: (1) surgical phases from the same or other procedures, (2) surgical steps from different procedure types, and (3) intraoperative time progression. Using datasets from four robotic-assisted procedures (sleeve gastrectomy, hysterectomy, cholecystectomy, and radical prostatectomy), we simulate real-world annotation scarcity by varying the proportion of available step annotations ( $\alpha$ $\in$ 0.25, 0.5, 0.75, 1.0). We benchmark the performance of a 2D CNN model trained with and without weak label pre-training.

Results: Pre-training with surgical phase labels-particularly from the same procedure type (PHASE-WITHIN)-consistently improved step recognition performance, with gains up to 6.4 f1-score points over standard ImageNet-based models under limited annotation conditions ( $\alpha$ = 0.25 on SLG). Cross-procedure step pre-training was beneficial for some procedures, and time-based labels provided moderate gains depending on procedure structure. Label efficiency analysis shows the baseline model would require labeling an additional 30-60 videos at $\alpha$ = 0.25 to match the performance achieved by the best weak-pretraining strategy across procedures.

Conclusion: Weakly supervised pre-training offers a practical strategy to improve surgical step recognition when annotated data is scarce. This approach can support scalable feedback and assessment in surgical training workflows where comprehensive annotations are infeasible.

Purpose: Radiofrequency ablation for liver cancer has advanced rapidly. For accurate ultrasound-guided soft-tissue puncture surgery, it is necessary to fuse intraoperative ultrasound images with preoperative computed tomography images. However, the conventional method is difficult to estimate and fuse images accurately. To address this issue, the present study proposes an algorithm for registering cross-source point clouds based on not surface but the geometric features of the vascular point cloud.

Methods: We developed a fusion system that performs cross-source point cloud registration between ultrasound and computed tomography images, extracting the node, skeleton, and geomatic feature of the vascular point cloud. The system completes the fusion process in an average of 14.5 s after acquiring the vascular point clouds via ultrasound.

Results: The experiments were conducted to fuse liver images by the dummy model and the healthy participants, respectively. The results show the proposed method achieved a registration error within 1.4 mm and decreased the target registration error significantly compared to other methods in a liver dummy model registration experiment. Furthermore, the proposed method achieved the averaged RMSE within 2.23 mm in a human liver vascular skeleton.

Conclusion: The study concluded that because the registration method using vascular feature point cloud could realize the rapid and accurate fusion between ultrasound and computed tomography images, the method is useful to apply the real puncture surgery for radiofrequency ablation for liver. In future work, we will evaluate the proposed method by the patients.

Purpose: For diffusion MRI (dMRI) parameter estimation, machine-learning approaches have shown promising results so far including the synthetic Q-space learning (synQSL) based on regressor training with only synthetic data. In this study, we aimed at the development of a new method named synthetic X-Q space learning (synXQSL) to improve robustness and investigated the basic characteristics.

Methods: For training data, local parameter patterns of 3 × 3 voxels were synthesized by a linear combination of six bases, in which parameters are estimated at the center voxel. We prepared three types of local patterns by choosing the number of bases: flat, linear and quadratic. Then, at each location of 3 × 3 voxels, signal values of the diffusion-weighted image were computed by the signal model equation for diffusional kurtosis imaging and Rician noise simulation. The multi-layer perceptron was used for parameter estimation and was trained for each parameter with various noise levels. The level is controlled by a noise ratio defined as a fraction of the standard deviation in the Rician noise distribution normalized by the average b = 0 signal values. Experiments for visual and quantitative validation were performed with synthetic data, a digital phantom and clinical breast datasets in comparison with the previous methods.

Results: By using synthetic datasets, synXQSL outperformed synQSL in the parameter estimation of noisy data sets. Through the digital phantom experiments, the combination of synXQSL bases yields different results and a quadratic pattern could be the reasonable choice. The clinical data experiments indicate that synXQSL suppresses noises in estimated parameter maps and consequently brings higher contrast.

Conclusion: The basic characteristics of synXQSL were investigated by using various types of datasets. The results indicate that synXQSL with the appropriate choice of bases in training data synthesis has the potential to improve dMRI parameters in noisy datasets.

Purpose: Carotid plaque is an early manifestation of carotid atherosclerosis, and its accurate segmentation helps to assess cardiovascular disease risk. However, existing carotid artery segmentation algorithms are difficult to accurately capture the structural features of morphologically diverse plaques and lack effective utilization of multilayer features.

Methods: In order to solve the above problems, this paper proposes a multi-scale hybrid attention hierarchical fusion U-network structure (MHAHF-UNet) for segmenting ambiguous plaques in carotid artery images in order to improve the segmentation accuracy for complex structured images. The structure firstly introduces the median-enhanced orthogonal convolution module (MEOConv), which not only effectively suppresses the noise interference in ultrasound images, but also maintains the ability to perceive multi-scale features by combining the median-enhanced ternary channel mechanism and the depth-orthogonal convolution space mechanism. Secondly, it adopts the multi-fusion group convolutional gating module, which realizes the effective integration of shallow detailed features and deep semantic features through the adaptive control strategy of group convolution, and is able to flexibly regulate the transfer weights of features at different levels.

Results: Experiments show that the MHAHF-UNet model achieves a Dice coefficient of $82.46\pm 0.31\%$ and an IOU of $71.45\pm 0.37\%$ in the carotid artery segmentation task.

Conclusion: The model is expected to provide strong support for the prevention and treatment of cardiovascular diseases.

Purpose: The incidence of acetabular and pelvic fractures is rising significantly. Pelvic ring fractures rank as the sixth most common fractures in adults, with the majority occurring in the elderly. Due to complications related to surgical approaches, with rates of up to 31%, there is an increasing demand for minimally invasive surgical techniques. Augmented Reality (AR) has the potential to facilitate spatial orientation by a sophisticated user interface. The aim of this study was to develop an AR-based, radiation-free navigation system for pelvic fractures.

Methods: The Microsoft® HoloLens 2 was used as the AR headset. The Unity® game engine was used for programming. Pelvic models from Sawbones® served as the model. Segmentation was performed using Slicer3D by Slicer Corporation. The symphysis and both anterior superior iliac spines were defined as anatomical reference points. Ten pelvic models were used for testing. A preoperatively defined drill trajectory was displayed to the surgeon. A total of 20 S1 screws and 19 S2 screws were placed using only AR navigation without visual access to the pelvic model. Screw placement was controlled using CT.

Results: The matching process took an average of 3 min and 28 s. 18 out of 20 (90%) S1 screws and 3 out of 20 (15%) S2 screws were placed correctly. In most cases, no perforation occurred. The mean procedure time was 7 min for S1 screws and 5 min for S2 screws.

Conclusion: Proper drilling was achieved by displaying the trajectories via AR, particularly for S1 screws, where a slightly wider drilling corridor was aimed for compared to S2 screws. No registration scan was necessary with our matching method. No intraoperative radiation was required.

Purpose: Augmented reality (AR) enhances surgical navigation by superimposing visible anatomical structures with three-dimensional virtual models using head-mounted displays (HMDs). In particular, interventions such as open liver surgery can benefit from AR navigation, as it aids in identifying and distinguishing tumors and risk structures. However, there is a lack of automatic and markerless methods that are robust against real-world challenges, such as partial occlusion and organ motion.

Methods: We introduce a novel multi-device approach for automatic live navigation in open liver surgery that enhances the visualization and interaction capabilities of a HoloLens 2 HMD through precise and reliable registration using an Intel RealSense RGB-D camera. The intraoperative RGB-D segmentation and the preoperative CT data are utilized to register a virtual liver model to the target anatomy. An AR-prompted Segment Anything Model (SAM) enables robust segmentation of the liver in situ without the need for additional training data. To mitigate algorithmic latency, Double Exponential Smoothing (DES) is applied to forecast registration results.

Results: We conducted a phantom study for open liver surgery, investigating various scenarios of liver motion, viewpoints, and occlusion. The mean registration errors (8.31 mm-18.78 mm TRE) are comparable to those reported in prior work, while our approach demonstrates high success rates even for high occlusion factors and strong motion. Using forecasting, we bypassed the algorithmic latency of 79.8 ms per frame, with median forecasting errors below 2 mms and 1.5 degrees between the quaternions.

Conclusion: To our knowledge, this is the first work to approach markerless in situ visualization by combining a multi-device method with forecasting and a foundation model for segmentation and tracking. This enables a more reliable and precise AR registration of surgical targets with low latency. Our approach can be applied to other surgical applications and AR hardware with minimal effort.

Purpose: In recent years, simulator-based training has gained attention as a safe and effective approach for surgical education. In orthopedic surgery, simulators for arthroscopic procedures, which require extensive practice to master, have been developed and shown to benefit clinical performance. This study aimed to evaluate the effectiveness of 2 types of arthroscopy simulators in medical education by assessing medical students without prior surgical training.

Methods: We prospectively and randomly evaluated the effectiveness and satisfaction of arthroscopy simulators using 2 devices: a high-fidelity virtual reality simulator (ArthroSim) and a low-fidelity simulator (AZBOTS). To assess simulator effectiveness, 28 first-year medical students received 90 min of training on either device. Their knee arthroscopy skills, including tasks such as visualizing and probing the medial compartment and cruciate ligaments, were assessed using ArthroSim. Skill acquisition was measured by procedure completion time and the Arthroscopic Surgery Skill Evaluation Tool Global Rating Scale. Additionally, to assess trainee satisfaction, 109 fifth-year medical students completed a 7-point Likert scale questionnaire evaluating spatial judgment, hand-eye coordination, camera navigation, instrument handling, and the overall knee arthroscopy training experience.

Results: No significant differences in skill acquisition were observed between the 2 simulator groups. Likewise, there were no significant differences in total questionnaire scores or in spatial judgment, hand-eye coordination, and camera navigation. However, participants using the high-fidelity simulator reported significantly greater satisfaction with instrument handling (p = 0.024) and the overall knee arthroscopy training experience (p = 0.033).

Conclusions: Although skill acquisition did not differ significantly between the high- and low-fidelity simulators after a single training session, the high-fidelity simulator markedly improved trainee satisfaction, especially in instrument handling and perceived quality of the training experience. These findings suggest that simulator fidelity enhances the educational value of the arthroscopic training for novices emphasizing the role of realistic simulation in early surgical education.

Purpose: NousNav is a low-cost, open-source neuronavigation platform built to address the high costs and resource limitations that hinder access to advanced neurosurgical technologies in low-resource settings. The low-cost and accessibility of the system is made possible using consumer-grade optical tracking and open-source software packages. This study aims to assess the performance of these core enabling technologies by quantifying their spatial accuracy and comparing it to a commercial gold standard.

Methods: A series of experiments were conducted to evaluate the capabilities of the selected hardware and registration infrastructure utilized in NousNav. Each component was tested both in a simulated bench-top environment and clinically across four brain tumor resection cases.

Results: The Optitrack Duo tracker used by NousNav was found to have a mean localization error of 0.8mm (SD 0.4mm). In bench-top phantom testing, NousNav had an average target registration error of 5.0mm (SD 2.3mm) following patient registration. Clinical evaluations revealed a mean distance of 4.2mm (SD 1.5mm) between points reported by NousNav versus those obtained using a commercial neuronavigation system.

Conclusion: These experiments highlight the role of baseline camera tracking performance, tracked instrument calibration, and patient positioning on the spatial performance of NousNav. They also provide an essential benchmark assessment of the system to help inform future clinical use-cases and direct ongoing system development.

Purpose: Distinguishing idiopathic normal pressure hydrocephalus (iNPH) from progressive supranuclear palsy (PSP) presents a clinical challenge due to overlapping clinical symptoms such as gait disturbances and cognitive decline. This study presents a novel multi-scale deep learning framework that integrates global and local magnetic resonance imaging (MRI) features using a mixture of experts (MoE) mechanism, enhancing diagnostic accuracy and minimizing interobserver variability.

Methods: The proposed framework combines a 3D convolutional neural network (CNN) for capturing global volumetric features with a 2.5D recurrent CNN focusing on disease-specific regions of interest (ROIs), including the lateral ventricles, high convexity sulci, midbrain, and Sylvian fissures. The MoE mechanism dynamically weights global and local features, optimizing the classification process. Model performance was assessed using stratified fivefold cross-validation on T1-weighted MRI from 118 patients (53 iNPH, 65 PSP) to ensure balanced class distributions across training folds.

Results: The MoE model using ResNet-34 achieved an accuracy of 0.983 (95% CI 0.875-1.000), a recall of 0.985 (95% CI 0.750-1.000), a precision of 0.986 (95% CI 0.769-1.000), and an area under the curve (AUC) of 1.000 (95% CI 1.000-1.000), outperforming traditional morphological markers and single-branch deep learning models. The MoE mechanism allowed adaptive weighting of global and local features, contributing to both improved robustness and interpretability. Grad-CAM visualizations highlighted disease-specific regions, demonstrating that the model focused on relevant features in both successful and failure modes of the 3D CNN expert for iNPH and PSP.

Conclusion: The dynamic integration of global and local MRI features through the MoE framework offers a powerful, robust, and interpretable tool for differentiating iNPH from PSP. This approach reduces reliance on subjective visual assessments and has the potential for broader clinical application through dataset expansion and multicenter validation.