The 2-D ultrasonography has shown great potential for evaluating periodontal structures due to its portability, flexibility, and nonionizing nature. However, for clinical applications, 3-D ultrasound imaging of oral anatomy is essential. This study presents the development of a freehand 3-D intraoral imaging system using the high-frequency ultrasound for reconstructing the 3-D anatomical structures of teeth. The proposed hand-held system integrates an in-house designed ultrasound transducer with positioning markers and four optical cameras to acquire the 2-D ultrasound images along with corresponding 3-D tracking poses. A forward-mapping algorithm was employed for 3-D data fusion to enable the real-time imaging. Both dimensional measurements and 3-D visualized difference colormaps were generated for a total of 24 teeth from maxillary and mandibular phantoms to compare the proposed system with an optical scanner and cone-beam computed tomography (CBCT). The mean distance differences ranged from -0.13 to 0.18 mm versus the optical scanner and from -0.18 to 0.15 mm versus CBCT. The overall mean absolute difference (MAD) across all dimensional measurements of two raters was 0.17 mm for the optical scanner and 0.16 mm for CBCT, respectively. The correlation coefficient with both reference modalities exceeded 0.98. The average model-to-model distance between the 3-D volumes acquired from the proposed system and the corresponding reference volumes was approximately 0.20 mm for the optical scanner and 0.28 mm for CBCT. The proposed 3-D intraoral imaging system demonstrated the feasibility of achieving high-quality 3-D intraoral imaging in real time, showing the potential to significantly accelerate the clinical adoption of the ultrasound technology in dentistry.

Hypoxic Ischemic Encephalopathy (HIE) represents a brain dysfunction, affecting approximately 1 to 5 per 1000 full-term neonates. The precise delineation and segmentation of HIE-related lesions in neonatal brain Magnetic Resonance Images (MRI) are pivotal in advancing outcome predictions, identifying patients at high risk, elucidating neurological manifestations, and assessing treatment efficacies. Despite its importance, the development of algorithms for segmenting HIE lesions from MRI volumes has been impeded by data scarcity. Addressing this critical gap, we organized the first BONBID-HIE challenge with diffusion MRI data (Apparent Diffusion Coefficient (ADC) maps) for HIE lesion segmentation, in conjunction with the MICCAI 2023. Totally 14 algorithms were submitted, employing a gamut of cutting-edge automatic machine-learning-based segmentation algorithms. Our comprehensive analysis of HIE lesion segmentation and submitted algorithms facilitates an in-depth evaluation of the current technological zenith, outlines directions for future advancements, and highlights persistent hurdles. To foster ongoing research and benchmarking, the annotated HIE dataset, developed algorithm dockers, and unified evaluation codes are accessible through a dedicated online platform (https://bonbid-hie2023.grand-challenge.org).

PURPOSE:To present comprehensive development and evaluation methodologies for a generalizable deep learning (DL)-driven autocontouring model of standard pelvic organs-at-risk (OARs) in MRI-planned cervical brachytherapy. MATERIALS AND METHODS:A curated dataset of 200 3D-MRIs (85% training/validation, 15% testing) including multiple applicator types, varying treated anatomies, and manual contours of OARs (bladder, rectum, sigmoid, small bowel) by 3 physicians was utilized to develop an nnU-Net-based autocontouring model. Iterative tuning was conducted to determine the optimal hyperparameters and enhance evaluation metrics. Model performance was assessed using quantitative metrics, like geometric (e.g., Dice Coefficient (DC) and Hausdorff Distance 95th Percentile (HD95)) and dosimetric (dose-volume histograms (DVHs), dose differences (ΔD2cc)), and then correlated with qualitative physician-review (modified Turing and Likert tests). RESULTS:Geometric metrics were best for bladder (e.g., mean ± SD DC|HD95(mm) 0.93 ± 0.02|2.26 ± 1.07) with greater variability exhibited for small bowel (0.62 ± 0.16|24.90 ± 14.36). Dosimetric comparisons of manual vs predicted contours showed high agreement in DVHs, with mean ΔD2cc <0.60 Gy EQD23 across all OARs. Model performance was consistent, irrespective of applicator type, OAR volume, or contourer. Quantitative scores in support of DLM were not always associated with as favorable qualitative results, yet physician-review showed clinical acceptability (80% for bladder and rectum). CONCLUSION:The DL-based autocontouring model, trained on a heterogeneous in-house dataset, demonstrates clinical acceptability for OARs as determined by comprehensive evaluation. It also shows promise for translatability to target contouring, and adaptability to other gynecological (noncervix) brachytherapy applications. Differences in qualitative and quantitative results exist; directionality and magnitude should be considered in clinical usability assessments of brachytherapy autocontouring models.

Objectives To evaluate the accuracy, efficiency, and reliability of manual, artificial intelligence (AI)-driven, and AI-assisted tracing methods for locating the mandibular canal (MC) on cone-beam computed tomography (CBCT) across clinical scenarios and evaluators. Methods Phase 1 established a calibration reference standard using a dry human mandible. In Phase 2, five evaluators assessed ten CBCT scans using manual and AI-assisted methods, while one expert performed all methods. The dataset included a range of anatomical variations and alterations. MC tracings were evaluated using Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD95), and boundary precision differences was analyzed with a linear mixed-effects model. Accuracy, efficiency, and agreement were assessed. Results A total of 292 MC tracings were evaluated. Manual tracing showed excellent intra-evaluator consistency (DSC: 0.99, HD95: 1.11) but low inter-evaluator agreement (ICC: 0.30). AI-assisted methods demonstrated near-perfect intra-evaluator reliability (DSC: 0.81, HD95: 2.2) and good inter-evaluator agreement (ICC: 0.72, DSC: 0.91, HD95: 1.23). Manual tracing showed the highest accuracy (DSC: 0.98) and was the slowest (6.2 min). The AI-driven method showed strong accuracy (DSC: 0.87) and was the fastest (13-20 s). The AI-assisted method balanced speed (4.2 min) and accuracy (DSC: 0.83), with superior boundary precision compared to manual (p = 3×10-4) and AI-driven methods (p = 1.0). Conclusion Manual tracing remains the most accurate method, time-intensive, and variable across evaluators. AI-driven tracing improved efficiency but was less reliable in complex cases. AI-assisted tracing balanced accuracy and efficiency, improved boundary precision, and reduced variability, supporting its role in clinical decision-making. Clinical Relevance The AI-assisted mandibular canal tracing method provides an effective balance between accuracy and efficiency. This reduces inter-user variability and improves reliability across users, even in complex anatomical cases.

OBJECTIVE:Accurate assessment of left ventricular (LV) function using three-dimensional echocardiography (3-DE) remains limited by suboptimal image quality and restricted field of view. This study proposes a robotic-arm-assisted acquisition protocol combined with a wavelet-based multi-apical view fusion approach to enhance LV image quality in 3-DE. METHODS:Volunteer scans were acquired using a UR10e robotic arm integrated with a Philips EPIQ 7C ultrasound system to ensure consistent multi-apical 3-DE acquisition. Echocardiographic volumes were converted to NRRD format using 3-D Slicer for visualization and verification of spatial and temporal alignment. Two-view and three-view apical datasets were fused using a wavelet-based approach. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were computed from matched 2-D slices in the end-diastolic phase, and qualitative image assessment was conducted by expert raters using blinded scoring of image clarity, myocardial border continuity and diagnostic confidence. RESULTS:Wavelet-based fusion significantly improved image quality compared to single-view 3-DE, with increased SNR (9.36 ± 5.03 vs. 7.09 ± 4.44, p < 0.0001) and CNR (1.68 ± 0.54 vs. 1.49 ± 0.57, p = 0.0020). Three-view fusion provided additional quantitative improvement over 2-view fusion. Inter-rater agreement on visual assessment confirmed that fused images were consistently rated as equal or superior in quality, with substantial agreement across all scoring categories. CONCLUSION:Wavelet-based fusion of multi-apical 3-DE images acquired with robotic arm assistance significantly enhances image quality for LV assessment, improving both quantitative metrics and visual interpretability, practically with the 3-view fusion. The use of the robotic arm played a key role in ensuring standardized and reproducible probe positioning, which is essential for successful image alignment and fusion. This approach demonstrates the potential to improve the reliability and diagnostic value of 3-DE, and future work should explore incorporating additional views and deep learning methods to further advance robotic-assisted cardiac imaging.

Segmenting ischemic stroke lesions on T1-weighted (T1w) MRI acquired across different scanners and protocols without intensity standardization is difficult because lesions are subtle and share intensity characteristics with cerebrospinal fluid. Standard deep learning architectures trained across multiple centers plateau around Dice 0.66, with acute lesions (≤ 7 days post-stroke) performing substantially worse due to severe sample scarcity. We combine a MedNeXt-L (k=5) backbone with on-the-fly 3D CarveMix augmentation that pastes real lesion patches into healthy brain regions during training. By generating synthetic lesion placements dynamically within each fold with subject-level split isolation, the model sees more diverse lesion patterns without requiring pre-generated copies on disk. We evaluate on 1,453 native T1w scans from 55 clinical centers in the ISLES 2026 challenge. Our method achieves a mean 5-fold cross-validation Dice of 0.648 at 500 epochs, a +0.018 improvement over the MedNeXt-L backbone at a matched training budget (0.630)

Three-dimensional (3D) echocardiography is an inherently noisy modality, but allows for capture of temporally complex cardiac motions. Existing segmentation methods often treat cardiac frames independently, resulting in inconsistent delineations across the cardiac cycle, reduced segmentation accuracy, and, in turn, impacting the estimation of key clinical metrics. To address this, our work introduces temporal positional embeddings based on the cardiac cycle to improve 3D echocardiography segmentation. By encoding cardiac phase information into sinusoidal functions, we inject temporal embeddings into the bottleneck and decoder of a U-Net. Our approach (TU-Net) outperforms state-of-the-art models, including nnU-Net and Transformer baselines, UNETR and SwinUNETR. TU-Net achieved a Dice score of 84.7% along with improved temporal consistency, evaluated over two test cases with 18 and 16 frames across the cardiac cycle. Testing this on a lightweight U-Net, which is both efficient and suitable for clinical settings, demonstrates the significance of temporal information in enhancing segmentation quality without complex models. These results highlight the potential for further improvements, not only in 3D echocardiography but also in other dynamic medical imaging modalities.

Pediatric cardiac disorders include an extensive range of heart conditions in infants, children and carried over to adolescents in certain cases. These disorders may be congenital or acquired and can vary in impacting the life of the pediatric subject which requires complex surgical procedures and certain cases that do not require medical intervention. Congenital heart disease (CHD) is present during birth and acquired cardiac diseases are developed after birth predominantly due to autoimmune responses or infections. Various acquisition techniques help in visualizing the heart, identify disorders and help the physicians to plan for operative procedures. Pediatric Cardiac Screening is one of the crucial techniques to record cardiac activity which has difficulties in acquisition as children tend to move during the procedure. The data obtained from these modalities may suffer from various artifacts which makes the diagnosis difficult for the clinicians. To make the diagnosis easier and artifact free, artificial intelligence plays a vital role. Artificial Intelligence (AI)-based techniques from traditional machine learning (ML) to deep learning (DL) techniques for classification and segmentation of pediatric cardiac signals and images are systematically reviewed. DL based studies have become a choice of research in pediatric healthcare. Support Vector Machines with linear kernels are the most commonly used ML based classifiers in the reviewed papers. DL methods use Convolutional Neural Networks (CNN) as the primary classifier and U-Net architectures are preferred for segmentation studies in the reviewed papers. There are a very few surveys available to present the diagnostics related to pediatric cardiac disorders and this paper tries to bring out the challenges along with the traditional machine learning and emerging deep learning techniques implemented in classification and segmentation of the pediatric cardiac disorders.

Segmentation of the right ventricle (RV) in magnetic resonance imaging (MRI) sequences is critical for assessing RV function. However, manual segmentation involves processing hundreds of images per patient, making it a tedious and timeconsuming process. Recently, deep convolutional neural networks have emerged as an effective solution for automating RV segmentation in MRI sequences, substantially reducing manual workload. Accurate segmentation of the RV is crucial for reliable clinical applications. In this study, we demonstrate that transfer learning using a pre-trained segmentation model from the Medical Open Network for Artificial Intelligence (MONAI) Model Zoo significantly improves segmentation accuracy, as measured by the Dice similarity coefficient (DSC) and $\mathbf{9 5}^{\text {th }}$ percentile Hausdorff distance (HD95) scores, compared to manual annotations from medical experts. Our approach increased DSC-based segmentation accuracy from 74.93 % (pre-trained MONAI Zoo model) and 83.15 % (same architecture trained on our data) to 84.91 % on 1,994 test images acquired from seven patients. Furthermore, it outperformed a state-of-the-art self-configuring network, nnU-Net, which achieved an accuracy of 81.98 % on the same dataset. This study demonstrates the effectiveness of transfer learning in improving segmentation accuracy for the proposed task.

Echocardiography remains a widely used imaging modality for the evaluation of cardiac structure and function. Despite its diagnostic value, conventional manual scanning techniques require sonographers to maintain repetitive postures and apply sustained pressure over extended periods, increasing the risk of work-related musculoskeletal disorders. In recent years, collaborative robots, or cobots, have emerged as a promising solution for applications requiring robots to operate safely alongside human operators. Modern cobot systems are often equipped with integrated force and torque sensors, enabling precise control of contact forces during patient scanning to ensure both safety and comfort. This study investigates the feasibility of a robotic-assisted echocardiography system in a clinical setting, focusing on its potential to reduce physical strain on sonographers while maintaining diagnostic image quality. A patient-based evaluation over 24 participants was conducted to assess system performance, force control accuracy, and image quality compared to conventional manual scanning. The findings aim to provide insights into the integration of robotic assistance in echocardiography workflows, with implications for improving operator ergonomics, patient safety, and imaging quality.

Background: Echocardiography is crucial to understanding cardiac function in the Intensive Care Unit (ICU), often by measuring the left ventricular ejection fraction (LVEF). Traditionally, measures of LVEF are completed as part of a comprehensive examination by an expert sonographer or cardiologist, but front-line practitioners increasingly perform focused point-of-care estimates of LVEF while managing life-threatening illness. The two main echocardiographic windows used to grossly estimate LVEF are parasternal and apical windows. Artificial intelligence (AI) algorithms have recently been developed to assist non-experts in obtaining and interpreting point-of-care ultrasound (POCUS) echo images. We tested the feasibility, accuracy and reliability of novice users estimating LVEF using POCUS-AI echo. Methods: A total of 30 novice users (most never holding an ultrasound probe before) received 2 h of instruction, then scanned ICU patients (10 patients, 80 scans) using the Exo Iris POCUS probe with AI guidance tool. They were permitted up to 5 min to attempt parasternal long axis (PLAX) and apical 4 chamber (A4C) views. AI-reported LVEF results from these scans were compared to gold-standard LVEF obtained by an expert echo sonographer. To further assess accuracy, this sonographer also scanned another 65 patients using Exo Iris POCUS-AI vs. conventional protocol. Results: Novices obtained images sufficient to estimate LVEF in 96% of patients in <5 min. Novices obtained PLAX views significantly faster than A4C (1.5 min vs. 2.3 min). Inter-rater reliability of LVEF estimation was very high (ICC 0.88–0.94) whether images were obtained by novices or experts. In n = 65 patients, POCUS-AI LVEF was highly specific for a decreased LVEF ≤ 40% (SP = 90% for PLAX) but only moderately sensitive (SN = 56–70%). Conclusions: Estimating cardiac LVEF from AI-enhanced POCUS is highly feasible even for novices in ICU settings, particularly using the PLAX view. POCUS-AI LVEF results were highly consistent whether performed by novice or expert. When AI detected a decreased LVEF, it was highly accurate, although a normal LVEF reported by POCUS-AI was not necessarily reassuring. This POCUS-AI tool could be clinically useful to rapidly confirm a suspected low LVEF in an ICU patient. Further improvements to sensitivity for low LVEF are needed.

Periodontal disease is a leading cause of tooth loss and is linked to systemic conditions such as endocarditis, diabetes, cardiovascular disease, and osteoporosis. Intraoral ultrasound (IUS) videos offer a non-invasive means for diagnosing periodontal structures, but existing segmentation methods rely on extensive manual annotations. We propose OralSAM, a one-shot video segmentation network inspired by the Segment Anything Model (SAM), which requires annotation from only a single frame. Our network integrates an adaptive feature correlation module to capture temporal dependencies and refine segmentation consistency across frames. Additionally, we introduce a self-prompting strategy based on optical flow, dynamically adjusting point prompts based on motion cues in consecutive frames to improve segmentation accuracy. To further enhance robustness, we incorporate a self-correction mechanism that refines mask embeddings adaptively, reducing propagation errors in intermediate frames. The combination of these components ensures effective generalization to unseen anatomical structures and improves temporal coherence in IUS videos. We evaluate OralSAM on both IUS and public datasets, demonstrating superior performance over state-of-the-art methods. Unlike conventional methods, our approach significantly reduces annotation effort while maintaining high segmentation accuracy. Our approach provides a scalable solution for real-time clinical applications, enabling more efficient and accurate periodontal disease assessment. Code is available at https://github.com/BioMedCom/OralSAM.

Cardiac magnetic resonance imaging (CMR) is considered the gold standard for assessing cardiac function. However, acquiring high-quality images typically requires patients to hold their breath during scanning. Free-breathing (FB) CMR serves as an alternative for patients who cannot hold their breath; however, it often introduces motion artifacts, degrading image quality and potentially affecting diagnostic accuracy. Although deep generative models have shown promise in correcting motion artifacts, ensuring confidence in the fidelity of reconstructed artifact-free images remains a significant concern. This study quantifies the impact of segmentation masks as guidance in diffusion models to enhance anatomical structure preservation during image-conditioned generation. To that end, a standard diffusion probabilistic model (DDPM) and a segmentation-guided DDPM are trained and evaluated on a public CMR dataset and further applied to restore FB CMR using local hospital data. Quantitative evaluations using signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) over the left and right ventricular regions demonstrate that the segmentation-guided approach produces higher-quality CMR and more accurately preserves anatomical structures compared to the standard DDPM.

Echocardiography is a non-invasive, non-ionizing, and cost-effective medical imaging modality that uses ultrasound waves to evaluate cardiac function. Left ventricle (LV) analysis is crucial for diagnosing cardiac diseases. Segmentation of the LV from echocardiography images is a challenging, time-consuming process requiring manual contouring from experts and is prone to inter-observer variability. Five different deep-learning models are evaluated for automatic LV segmentation from 3D echocardiography. The models were compared using overlap and distance metrics: Dice score, Jaccard index, and Hausdorff distance. Volumetric analysis was used to examine the accuracy of the predictions from the deep learning models against the expert-annotated ground truth volumes. The comparison between these models provides a foundation for further development of accurate and efficient automated LV segmentation methods, particularly approaches that can leverage the temporal consistency of echocardiography scans.

BACKGROUND:Artificial intelligence (AI) is transforming diagnostic imaging in dentistry. This systematic review evaluates existing literature on augmented intelligence in dentomaxillofacial radiology, focusing on its influence on human collaboration in interpreting dental imaging. STUDY DESIGN:A literature search across seven databases and gray literature was conducted. Studies evaluating clinician performance with AI-assistance were included, while reviews, surveys, and case reports were excluded. The QUADAS-2 tool assessed the risk of bias. RESULTS:Sixteen studies assessed the influence of AI on radiographic interpretation. AI-assisted caries detection consistently improved accuracy, sensitivity, and specificity. Detection of apical pathoses and jaw lesion segmentation improved accuracy, reducing diagnostic time. Cephalometric landmark identification showed increased accuracy, particularly for students. Soft tissue calcification detection improved accuracy, but sensitivity decreased. Overall, augmented intelligence enhanced interobserver agreement and reduced diagnostic variability, with general dentists and students showing the greatest gains. CONCLUSIONS:Augmented intelligence enhances dental radiographic interpretation by improving tasks, particularly for less experienced clinicians, and positively influences clinical decision-making. However, AI performance remains inconsistent in challenging cases involving complex pathoses or varied imaging conditions. While it complements rather than replaces clinicians, further validation of AI's generalizability and reliability using larger, diverse datasets is necessary.

This article will be removed in accordance with Elsevier’s Article Withdrawal Policy (https://www.elsevier.com/about/policies-and-standards/article-withdrawal).This article was removed at the request of the authors.The manuscript includes research conducted using a beta version of a software tool. This beta version does not permit its outputs to be published, and its use may also affect the reported results. The authors agree with the software company’s request and therefore support the removal.

Objectives To develop a deep learning (DL) algorithm to segment the adenoid hypertrophy (AH) area from Cone Beam Computed Tomography (CBCT) scans to aid in the early detection of enlarged adenoids and improve management of AH. Methods This retrospective study utilized CBCT scans, comprising oral radiologist-graded scans for training and validation, and a test dataset diagnosed by an Ear, Nose, and Throat (ENT) specialist using nasoendoscopy (NE), which served as the reference standard for external validation. Manual adenoid area segmentation was performed using 3D Slicer. A DL algorithm, based on convolutional neural networks, was developed to segment the naso- and oropharynx in CBCT images with and without AH. The Dice Similarity Coefficient and Intersection over Union were applied to assess segmentation accuracy. Results A total of 96 CBCT scans, distributed by AH grading, reflected at least 22,800 DICOM manually segmented files. Evaluator calibration was confirmed within the intraclass correlation coefficient (ICC) 0.90. Data augmentation was applied, maintaining the dataset distribution. Four nnU-Net-based segmentation models were tested: 2D, 3D Full-resolution (3D Fullres), 3D Low-resolution (3D Lowres), and 3D Cascade. The algorithm achieved 0.90 overall accuracy, a 0.90 Dice score, and 0.08 precision on the test dataset for adenoid area segmentation. Conclusions The trained nnU-Net model demonstrated excellent results in segmenting the AH region, (Dice score: 0.99) achieved for the combination of 3D Cascade and 3D Fullres models. When applied in available imaging, this DL integration with CBCT enhances early AH detection and streamlines referrals for timely treatment by medical teams.