Search Results (475)

Accurate stereo visual–inertial measurement is essential for mobile robots operating in Global Navigation Satellite System (GNSS)-denied and structurally complex environments. In stereo visual–inertial odometry (VIO), pose and trajectory outputs depend strongly on the point measurements delivered by the visual front end before sensor-fusion update. In sparse-texture but structurally regular scenes, tracked point features may exhibit poor persistence, uneven spatial distribution, and local tracking noise, even when informative line structures are present. Existing point–line VIO methods can improve positioning accuracy by introducing line landmarks or line residuals, but they usually modify the estimator state, measurement model, and Jacobian treatment. We present a scene-adaptive line-aware visual measurement conditioning method for stereo VIO front ends with point-measurement updates. The method uses 2-D image-line segments as lightweight structural priors and applies bounded normal-direction conditioning to reliable point measurements before a fixed-interface VIO back-end update. A sparse pruning safeguard removes only highly inconsistent long-lived tracks under strong structural support, while a scene-level confidence gate attenuates the intervention when line evidence is weak or unstable. The method is instantiated and evaluated in an S-MSCKF pipeline. On the reported EuRoC MAV sequences, it reduces the sequence-averaged absolute trajectory error (ATE) RMSE by approximately 13% relative to S-MSCKF, with 3–27% reductions on machine-hall sequences. On three real-world robot measurement sequences with an RTK-aided inertial reference, the mean Sim(2)-aligned planar position error decreases from 8.72 m to 7.31 m, and the mean yaw error decreases from $8.{02}^{\circ }$ to $6.{76}^{\circ }$; an additional scale-preserving SE(2) evaluation reveals sequence-dependent planar behavior and residual metric-scale sensitivity. Candidate-level stereo-consistency diagnostics show subpixel mean and 95th-percentile image-domain perturbations without systematic vertical-stereo bias, while the final reliability-weighted primary-view update is analytically bounded by approximately 0.221 pixels in the reported implementation. Runtime profiling reports an average front-end time of 33.34 ms on the tested CPU platform, close to the 33.3 ms frame period of the 30 Hz stereo input, although the $\mu +3\sigma$ runtime of 47.23 ms exceeds a strict frame-by-frame 30 Hz budget. These results suggest that line-aware front-end conditioning can improve visual measurement quality in structured stereo visual–inertial sensing without modifying the evaluated back-end interface. Full article

Reliable localization without GPS is a prerequisite for autonomous unmanned aerial vehicles (UAVs) operating inside warehouses, where a lightweight monocular camera paired with an inertial measurement unit (IMU) and onboard computer are the minimal sensing and processing an onboard platform can carry. This paper presents a tightly coupled monocular point-feature visual-inertial odometry (VIO) system for that setting, realized on a GTSAM fixed-lag factor graph with inverse-depth landmarks, on-manifold IMU preintegration, and an online loop-closure pose graph. The system is developed as the initial estimation stage of an autonomous stock-management UAV under development for indoor logistics warehouses. The decisive design element is the bootstrap: the metric, gravity-aligned initialization of a monocular estimator is well conditioned only under a translation-rich trajectory, a condition the near-zero-baseline pickup and takeoff transient that opens every indoor flight violates. Building on the visual-inertial alignment of VINS-Mono, we harden this step with a pre-bundle-adjust conditioning gate and a continuous-window initialization that refines the whole bootstrap window inside the smoother instead of freezing a single seed. On all eleven EuRoC MAV sequences, indoor flight tests recorded onboard a micro air vehicle in an industrial hall and two instrumented rooms, one fixed configuration per operating environment converges on every sequence, including three that otherwise diverge by tens to thousands of meters, and, driven by the same feature stream as locally run VINS-Mono and PL-VINS baselines, attains the better pure-odometry accuracy on nine of the eleven, with ATE RMSE of 0.12–0.37 m on the Machine Hall, a margin a paired signed-rank test confirms against VINS-Mono and leaves unconfirmed against PL-VINS at this sample size. We identify the stock fixed-lag marginalization as the principal consistency limitation and outline First-Estimates-Jacobian marginalization as the route to a more consistent estimator, establishing a characterized point-only baseline on one public benchmark as the starting point for subsequent on-platform work. Full article

Background: Single-time-point cerebrospinal fluid (CSF) volume proportion (pCSFV) has shown limited prognostic utility after out-of-hospital cardiac arrest (OHCA). We hypothesized that serial within-patient change in pCSFV (ΔpCSFV), rather than absolute values, would reflect progression of neurologic injury. Methods: This retrospective 72–96-h landmark cohort study included comatose adult OHCA survivors who underwent brain computed tomography (CT) within 6 h (early) and at 72–96 h (delayed) after return of spontaneous circulation. pCSFV was quantified using Hounsfield-unit-based volumetry. ΔpCSFV was calculated as delayed pCSFV minus early pCSFV. The primary outcome was poor 6-month neurologic outcome (Cerebral Performance Category 3–5). Results: Among 125 patients, 63 (50.4%) had poor neurologic outcomes. Early and delayed pCSFV did not differ significantly between outcome groups, whereas ΔpCSFV was lower in the poor-outcome group (p < 0.001). In the five-parameter multiple-imputation model, ΔpCSFV was associated with poor outcome (adjusted odds ratio [aOR] 0.684 per 1 percentage-point increase, 95% confidence interval [CI] 0.494–0.947; p = 0.022), with a consistent result in the complete-case sensitivity analysis (aOR 0.648, 95% CI 0.446–0.940; p = 0.022). However, this association was attenuated after additional adjustment for early pCSFV (aOR 0.728, 95% CI 0.462–1.146; p = 0.170). Adding ΔpCSFV to the baseline model did not significantly improve discrimination (area under the curve 0.962 vs. 0.955; p = 0.292). Conclusions: ΔpCSFV may provide complementary information on serial CSF-space changes but should not be interpreted as a baseline-independent marker. These findings apply only to patients who survived and underwent delayed CT at 72–96 h. The observed data-derived cutoff is exploratory and requires external validation before any clinical application. Full article

Urban anomalies such as rare facilities, unusual spatial configurations, and atypical functional patterns can provide valuable insights into urban dynamics and planning processes. However, graph-based anomaly discovery methods often suffer from instability, producing substantially different results across training runs and making the detected anomalies difficult to reproduce and interpret. In this paper, we use the term anomaly to denote automatically detected abnormal urban entities, while the term urban irregularity refers to their interpretation within the urban context. We present a consensus-driven framework for stable anomaly discovery using multi-relational point-of-interest (POI) graphs. Urban facilities are represented as nodes connected through multiple spatial and semantic relations, including geographic proximity, shared categories, shared facility types, and region-based associations. Four graph autoencoder architectures (GAE, ResGAE, VGAE, and SAGEAE) are employed to learn node representations, while reconstruction-, cluster-, neighborhood-, and relation-based anomaly scoring strategies are combined with multi-seed stability analysis to identify consensus anomalies. Experiments conducted on five large-scale cities (Baku, Turin, Vienna, Prague, and Kuala Lumpur) show that the proposed framework identifies recurring anomaly patterns across repeated runs and analytical configurations. Comparisons with representative anomaly detectors reveal partial but method-dependent overlap, while cross-city control experiments indicate that a subset of the detected anomalies exhibits non-random semantic and structural correspondence across different urban environments. Additional analyses suggest that consensus anomalies are frequently associated with semantically distinctive urban entities, including recreational areas, utility infrastructure, institutional facilities, specialized services, and cultural landmarks. Overall, the results indicate that stability-aware consensus provides a more reproducible and consistent basis for graph-based urban anomaly discovery and supports the interpretation of recurrent anomaly patterns in large-scale urban POI graphs, without requiring ground-truth anomaly labels. Full article

Flexible flatfoot (pes planus) alters lower-limb biomechanics and plantar-pressure distribution, raising the risk of pain and injury. Laboratory gait analysis with optical motion capture and force plates is the reference standard but is costly, space-constrained, and ecologically limited. We present the design, fabrication, and validation of a low-cost, sustainable smart insole for Internet-of-Things (IoT) remote body-load monitoring. The device pairs a dual-layer natural-rubber body—a silica-filled sponge–rubber upper for comfort and a carbon-black-reinforced solid outsole for durability—with four load cells per insole at high-pressure plantar landmarks, read through a 24-bit amplifier by an ESP32 that calibrates and streams left/right load over Wi-Fi to the ThingSpeak cloud, with a wrist-worn OLED for real-time feedback. Against reference weights in 25 participants, the system measured total body weight with a mean absolute error of 2.94%, a maximum error of 4.18%, and an RMSE of 1.94 kg (Pearson r = 0.99); the residual was an almost purely systematic proportional bias (slope 0.966, R² = 0.98) removable by a single in-sample scalar recalibration. In 30 adults (15 normal-arch; 15 flexible flatfoot), spatiotemporal gait parameters were compared while both groups wore the smart insole. Forward-progression parameters, including step length, stride length, and walking velocity, did not differ significantly between groups during comfortable walking (all p > 0.18). The flatfoot group showed a wider mediolateral base—greater stance width during standing (+14%, p = 0.008) and step width during comfortable walking (+23%, p = 0.040, uncorrected). After correction for multiple comparisons, only the reduction in fast-walking cadence remained statistically significant. A sustainably sourced, affordable smart insole can thus deliver clinically meaningful remote body-load monitoring. The findings also point to a dissociation: forward propulsion was comparable between the groups while the insole was worn, whereas the mediolateral base remained wider in flatfoot. Controlled trials pairing orthotic support with active gait retraining are therefore warranted. Full article

Background/Objectives: Glucose variability (GV) and the lactate-to-albumin ratio (LAR) have been associated with adverse outcomes in critically ill patients, but their incremental prognostic contribution beyond established severity assessment remains uncertain. This study evaluated the associations of GV and LAR with subsequent mortality and their incremental value beyond APACHE II in a conditional 48 h landmark cohort. Methods: This single-center retrospective cohort study included 384 critically ill adults who were alive and remained in the ICU through 48 h and had sufficient glucose, lactate, and albumin measurements. GV was quantified using the coefficient of variation (CV) from six glucose measurements closest to 0, 8, 16, 24, 32, and 48 h. The primary outcome was all-cause mortality during the 7 days following the 48 h landmark. Firth penalized logistic regression was used for prognostic modeling. Incremental performance was assessed sequentially for APACHE II, APACHE II + LAR, and APACHE II + LAR + GV using discrimination, calibration, Brier score, bootstrap internal validation, and decision-curve analysis. Sensitivity analyses adjusted for mean glycemia and restricted predictor information to the first 24 h. Results: Eighty-eight patients (22.9%) died during the 7-day post-landmark period. APACHE II (OR: 1.19 per point, 95% CI: 1.12–1.27; p < 0.001), LAR (OR: 1.97 per unit, 95% CI: 1.48–2.69; p < 0.001), and GV (OR: 1.07 per 1-percentage-point increase in CV, 95% CI: 1.02–1.11; p = 0.002) were independently associated with mortality. The AUC increased from 0.804 for APACHE II alone to 0.844 after addition of LAR and to 0.857 after further addition of GV. The additional AUC increase attributable to GV after LAR was modest (ΔAUC = 0.013, 95% CI: −0.004 to 0.030; p = 0.145), although model fit and Brier performance improved. The optimism-corrected C-index of the integrated model was 0.851. GV remained independently associated with mortality after adjustment for mean glucose (OR: 1.06, 95% CI: 1.02–1.11; p = 0.002) and in the temporally matched 24 h analysis (OR: 1.05, 95% CI: 1.01–1.09; p = 0.008). Conclusions: Among critically ill adults who survived to a 48 h landmark, LAR and GV provided prognostic information beyond APACHE II. Most of the incremental improvement in discrimination was attributable to LAR, whereas GV provided a smaller additional contribution that remained consistent across sensitivity analyses. These findings support further evaluation of LAR and GV as complementary prognostic markers, but external validation is required before clinical implementation. Full article

The shoe last is central to footwear design. Modern footwear CAD operates on parametric digital lasts, yet much last geometry—legacy collections and lasts that skilled last makers still sculpt by hand and copy by pantograph turning—exists only as physical models or as point-cloud scans lacking the structured parametric form that footwear CAD requires. This paper presents a complete template-based method for reconstructing a watertight parametric last from a segmented point cloud without intermediate triangulation. The only manual input is three landmark points—for which the system proposes standard positions—and the interactive confirmation of two boundary lines on the digitized last. From these, the method defines four feature points, a median plane, and a four-curve boundary network; all subsequent stages run without user interaction. A curvature-adaptive quadrilateral grid is constructed on the cloud by geodesic tracing and monitor-weighted area-orthogonality relaxation. A periodic Coons tube interpolates the grid and initializes the parameterization for a periodic tensor-product cubic B-spline surface fitted by penalized least squares with cyclic/open difference penalties, exact boundary interpolation, and toe-aware weighting. Cap surfaces close both collar and sole openings, and the model is exported as a watertight B-rep solid. Tests on sixteen industrial lasts using one fixed parameter set produced a mean one-sided deviation of 0.034 mm (RMS 0.058 mm) from the withheld industrial reference meshes in approximately 12 s per last. With synthetic noise at 50 dB SNR, the mean deviation increased by only 0.011 mm. A sampling-density study indicated near-second-order convergence before the control-net reaches an upper limit. The resulting solids import directly into CAD systems and support re-lasting, footwear design, and customization. Full article

Background/Objectives: Sodium–glucose cotransporter-2 (SGLT2) inhibitors improve cardiorenal outcomes in patients with diabetes, heart failure, and chronic kidney disease. Their association with perioperative outcomes after major cardiac procedures remains uncertain. Methods: In this retrospective cohort study using the TriNetX U.S. Collaborative Network, adults (≥18 years) undergoing cardiac procedures were identified. Patients with documented SGLT2 inhibitor use within 1 year before the index procedure were compared with non-users. Risk ratios (RR), risk differences, Kaplan–Meier survival curves, and Cox models were used for analysis. Results: Of 150,724 patients identified, 7526 propensity score-matched pairs were analyzed. The matched cohorts had a mean age of 65.8 ± 10.6 years and were predominantly male (70.5% male, 29.5% female). Preoperative SGLT2 inhibitor use was associated with significantly lower all-cause mortality at 30 days (1.3% vs. 2.3%; RR 0.566 [95% CI 0.443–0.722], p < 0.001), 90 days (2.2% vs. 3.4%; RR 0.643, p < 0.001), and 1 year (3.8% vs. 5.8%; RR 0.657, p < 0.001). MACE rates were similar between groups at all time points. Acute kidney injury at 30 days was lower with SGLT2 inhibitors (15.4% vs. 17.7%; RR 0.870, p < 0.001), whereas hospital admission rates were higher (76.4% vs. 68.7%; RR 1.111 [95% CI 1.089–1.133], p < 0.001). These associations were consistent in diabetes-only and obesity-only subgroups and in a 30-day landmark sensitivity analysis. Conclusions: In this large, contemporary, multi-institutional propensity score-matched cohort, preoperative SGLT2 inhibitor use was associated with reduced all-cause mortality and acute kidney injury after major cardiac procedures. These real-world findings support the conduct of dedicated randomized trials evaluating perioperative SGLT2 inhibitor therapy. Full article

Background and Objectives: Delayed cerebral ischemia (DCI) is a major determinant of poor outcome after aneurysmal subarachnoid hemorrhage (aSAH), yet early risk prediction remains difficult, particularly in sedated or ventilated intensive care patients. We evaluated whether a large language model (LLM) could predict DCI from phase-based dynamic clinical data, compared with intensive care specialists. Materials and Methods: In this single-center, retrospective study, 216 consecutive patients with aSAH were assessed at three predefined phases of accumulating clinical data (day 1; days 1 + 3; days 1 + 3 + 5). For each patient–phase, an LLM (ChatGPT, GPT-5.5 Thinking) and two blinded intensive care specialists predicted DCI risk using only the data available up to that time point. DCI was adjudicated by a blinded three-member panel. Discrimination was assessed by the area under the receiver operating characteristic curve (AUC), with non-inferiority defined a priori as Δ = 0.10. Calibration, decision-curve analysis, and reproducibility were also assessed. Results: Of 216 patients, 60 (27.8%) were DCI-positive and 15 were indeterminate; the primary sample comprised 201 patients. In the prespecified primary analysis, the LLM met the non-inferiority criterion relative to both specialists across all three phases (LLM AUC 0.703–0.747; specialists 0.712–0.764); however, in an equal-granularity sensitivity analysis, non-inferiority remained supported only in Phases 2 and 3 and was not demonstrated in Phase 1. Discrimination increased numerically as data accumulated (Phase 1 vs. 3, p = 0.072), an increase that was attenuated in a landmark-restricted analysis accounting for DCI-onset timing. The LLM showed the lowest false-reassurance rate (3.3–11.7%), reflecting a more cautious threshold rather than better discrimination. Confidence did not reliably indicate accuracy (~25% of high-confidence predictions were wrong); outputs were highly reproducible (Fleiss κ 0.887; intraclass correlation coefficient (ICC) 0.968). Conclusions: An LLM achieved DCI discrimination that was non-inferior to—but not better than—that of experienced specialists; its unreliable confidence scores support clinician-supervised rather than autonomous use. Full article

Background/Objectives: The prognostic nutritional index (PNI) has been associated with outcome in metastatic breast cancer, but studies in CDK4/6 inhibitor-treated patients have used cohort-derived thresholds and examined only the pretreatment value. We assessed the PNI as a continuous variable, compared it with inflammatory indices, and examined whether repeated measurement added information. Methods: We reviewed 192 consecutive patients with hormone receptor–positive, HER2-negative metastatic breast cancer treated with a CDK4/6 inhibitor and endocrine therapy at a single centre. We calculated the PNI, neutrophil-to-lymphocyte ratio (NLR), and systemic inflammation response index (SIRI) before treatment and recalculated them before cycles 2 and 3. The Cox models were adjusted for age, liver metastasis and line of therapy; no threshold was derived from these data. Results: The median follow-up was 44.9 months. Each one-point increase in the baseline PNI was independently associated with lower hazards of progression (HR 0.952, 95% CI 0.923–0.983) and death (HR 0.919, 95% CI 0.884–0.956), corresponding to HR 0.782 (95% CI 0.670–0.918) and HR 0.656 (95% CI 0.540–0.799) per five-point increase. For overall survival, in formal comparisons on identical patient sets, neither the NLR nor the SIRI added prognostic information to a model containing the PNI (likelihood ratio p = 0.383 and p = 0.335), whereas the PNI added information to models containing either ratio (p < 0.001 and p = 0.004). In exploratory landmark analyses, change in the PNI added little to the baseline value; apparent associations between increases in the NLR or SIRI by cycle 3 and overall survival did not persist after winsorisation of extreme change scores. Conclusions: Baseline PNI, analysed as a continuous variable and without a data-derived threshold, was independently associated with progression-free and overall survival in this cohort. For overall survival, neither baseline inflammatory ratio added detectable prognostic information beyond the index, and repeated measurement during treatment added little to the baseline value, although the confidence intervals do not exclude modest effects. Because all patients received a CDK4/6 inhibitor and no comparator arm was available, these findings support a prognostic rather than a predictive interpretation, and external validation is required before the index can inform clinical decisions. Full article

Background/Objective: Sex estimation from cranial remains is essential in forensic biological profile reconstruction, yet systematic comparisons of statistical classifiers under formally verified assumption conditions remain limited in Thai populations. This study aimed to develop and validate sex estimation models from craniofacial measurements in a Northeastern Thai skeletal sample using DFA, BLR, and SVM, and to examine whether distributional assumption violations influence classifier performance and generalizability under holdout validation. Methods: Eight linear craniometric parameters connecting five osteometric landmarks were measured in 300 adult skeletal specimens (150 males, 150 females) from the Unit of Human Bone Warehouse for Research, Khon Kaen University. Specimens were allocated into a training dataset (n = 200) and a holdout dataset (n = 100). Five multivariate model configurations were developed under direct-entry and stepwise parameter-entry strategies. Formal assumption diagnostics included Shapiro–Wilk normality testing, Levene’s test, Box’s M test, and variance inflation factor analysis. Results: All parameters demonstrated excellent interobserver reliability and significant sex-based differences (p < 0.01), with males recording larger mean values throughout. Apparent multivariate accuracy ranged from 78.0% to 85.5%. Holdout validation revealed consistent performance decline across all five configurations, with accuracy reductions of 3.0 to 9.5 percentage points and AUC reductions of 0.042 to 0.079. The stepwise DFA model showed the smallest overall accuracy decline, although this apparent stability reflected an uneven sex-specific trade-off, with a specificity decline offset by a sensitivity gain, while SVM and DFA direct-entry models showed the largest decline. Conclusions: Craniofacial measurements provide moderate discriminatory capacity for sex estimation in Northeastern Thai skeletal remains. Although univariate assumption violations were observed, DFA’s performance remained comparable to BLR and SVM under the present large, balanced sample; multivariate analysis or distributional assumption-free methods offer practical alternatives where violations are present. Apparent performance alone is insufficient to establish forensic reliability, and holdout validation should be regarded as a necessary component of sex estimation model development. Full article

Background: Accurate acetabular component anteversion during total hip arthroplasty (THA) remains challenging. We evaluated a preoperative computed tomography (CT)–based technique that uses patient-specific acetabular bony landmarks to position the cup. Methods: We prospectively enrolled 48 patients (50 hips) undergoing primary THA between September 2023 and January 2024. On preoperative CT, the anterior acetabular notch (AAN) was defined as the anterior pole, and the point opposite it across the acetabular center was defined as the posterior pole. Using the cup radius (r), the target-relative rotation (θ), and half the interpolar distance (R), the arc-length distances from each pole to the cup margin (a and b) were derived preoperatively with exact circular-geometry equations and reproduced intraoperatively. Postoperative CT anteversion was measured and converted to radiographic anteversion to determine outliers from the Lewinnek safe zone. Early clinical outcomes, including dislocation and modified Harris hip score (mHHS), were assessed at 3 months. Results: Mean postoperative CT anteversion was 25.1° (range, 16.0–35.0°), with a mean absolute error of 4.0° (range, 0–12.0°). Calculated radiographic anteversion was 16.8° (range, 10.0–24.0°), and all 50 hips were within the Lewinnek safe zone. At 3 months, mean mHHS was 92.4, and no dislocations were observed. Conclusions: This prospective proof-of-concept study suggests that CT-based acetabular bony landmarks can enable accurate cup positioning in standard primary THA. Further controlled studies are required to confirm reproducibility and clinical advantage. Full article

High-level autonomous driving requires an exceptionally accurate and robust perception of surrounding road environments to support reliable vehicle localization. While utilizing multiple LiDAR sensors expands the field of view and provides high-density point clouds, it introduces severe challenges, such as multi-LiDAR mutual interference (crosstalk) and significant computational overhead. This paper proposes a multi-stage front-end perception pipeline designed for robust high-altitude structural landmark extraction and crosstalk suppression to provide clean geometric reference points for downstream positioning systems. The proposed framework first employs a Binary Bayes Filter combined with a mesh-graph representation to segment stable overhead road traffic signs while minimizing environmental noise. To ensure real-time operation, a hybrid cascade consisting of 2D-grid spatial projection and voxelized 3D Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is introduced, drastically reducing point cloud density bottlenecks. Finally, a cluster-refinement stage is applied to filter out residual false positives induced by crosstalk. The system was validated using real-world driving data collected over an urban sequence (1.88 km) and a highway sequence (37 km). Experimental results demonstrate that the pipeline achieves high landmark tracking success rates of 79.9% and 98.5%, respectively, while suppressing crosstalk-induced false alarms down to 12.3%. Operating entirely on CPU threads with an average latency of 24.5 ms per frame, the framework satisfies real-time execution bounds for standard 10 Hz LiDAR setups, establishing a high-fidelity front-end capable of preventing tracking drift in autonomous vehicle localization. Full article

Objectives: Conventional overjet is widely used to describe sagittal interarch relationships, but its measurement depends on predefined landmarks and acquisition methods. This study aimed to describe a standardized three-dimensional method, termed 3D Clinical Overjet, for measuring the minimum sagittal distance between anterior dental surfaces in Class II malocclusion. Methods: This retrospective methodological study applied the proposed protocol to 42 consecutive growing patients with skeletal and dental Class II malocclusion. Digital intraoral scans obtained at baseline (T0) and immediately before mandibular advancement (T1) were analyzed by a single examiner. 3D Clinical Overjet was defined as the minimum linear distance, measured parallel to the occlusal plane, between the labial surface of any mandibular incisor and the palatal surface of the nearest maxillary incisor. No intra- or inter-examiner reliability assessment was performed because the study was designed to demonstrate feasibility rather than validate measurement properties. Associations with radiographic overjet were assessed using Spearman’s rank correlation; T0–T1 changes were analyzed using the Wilcoxon signed-rank test, and tooth-distribution changes were analyzed using a marginal homogeneity test. Results: 3D Clinical Overjet was obtained in all 42 subjects at T0 and in 41 at T1. Among the 41 subjects with paired observations, median 3D Clinical Overjet increased from 0.60 [0.10–2.00] mm at T0 to 1.70 [0.60–4.30] mm at T1 (W = 224, p = 0.054; rank-biserial correlation = 0.36). The maxillary tooth determining the minimum distance varied among patients and across time points. 3D Clinical Overjet was not significantly associated with radiographic overjet at T0 (Spearman ρ = 0.095, p = 0.558; n = 40). Conclusions: 3D Clinical Overjet is a feasible surface-based three-dimensional method for identifying the minimum sagittal distance across the anterior dentition without restricting measurement to a predefined incisor pair. In this sample, the tooth determining the minimum distance varied among patients and across time points, and 3D Clinical Overjet showed no significant association with radiographic overjet. These findings support the feasibility of the proposed geometric approach; its reproducibility, validity, and clinical relevance remain to be established. Full article

Background/Objectives: Deep learning models for cephalometric landmark detection report near-ceiling accuracy on single benchmarks, yet most are trained and tested on the same dataset. Whether the best in-domain architecture remains best out-of-domain has not been systematically quantified. Methods: Four architecture families (heatmap CNN, two-stage cascade, coordinate-regression Vision Transformer, pretrained ResNet-50) were each trained on two independently sourced datasets—ISBI 2015 (400 images, one device) and Aariz (1000 images, seven devices)—and evaluated on both, over their 19 shared landmarks in millimeters. All axes used three seeds (mean ± SD): the cross-dataset matrix, leave-one-device-out shift, balanced joint training, a pretrained-versus-scratch ablation, and a landmark-level breakdown. Results: In-domain mean radial error (MRE) was 2.47–3.04 mm (Aariz) and 4.12–5.55 mm (ISBI); cross-dataset error rose steeply (Generalization Drop—the relative increase in error out-of-domain—155–426%). The most accurate model in-domain (a from-scratch heatmap CNN, 2.47 mm) showed the largest drop, and every pretrained estimate fell below every from-scratch estimate (family means 248% vs. 380%): in-domain ranking did not predict cross-domain ranking. Leave-one-device-out revealed reproducible device-specific shift (held-out MRE 1.89–7.94 mm). Inter-observer variability was 0.53 mm, so cross-domain errors were 15–44× the human band. Balanced joint training reduced the cross-domain gap for all four architectures (both domains ≈ 2.1–3.5 mm) without harming in-domain accuracy. Pretraining more than halved cross-domain error from Aariz to ISBI (7.73 vs. 16.36 mm) but not in reverse, supporting the mechanism in one direction rather than uniformly. A-point, B-point, Nasion, and Menton all exceeded the 2 mm clinical threshold out-of-domain, including points localized to sub-millimeter accuracy in-domain. Conclusions: Single-dataset accuracy substantially overstates clinical generalizability, and the best in-domain architecture is not the most transferable, so in-domain leaderboards are an unreliable basis for selecting a model for clinical deployment; balanced multi-source training recovers most of the loss across the architectures tested. Full article