Search Results (219)

Soil salinity is an increasing constraint to crop production, yet little is known about the responses of culantro (Eryngium foetidum), a medicinal and culinary species of high economic value, to saline conditions. This study evaluated the effects of NaCl-induced salinity (0, 40, and 80 mM) on the growth, photosynthetic performance, pigment content, and vascular anatomy of E. foetidum cultivated in vitro. After 45 days, salinity significantly reduced shoot and root length, leaf number, chlorophyll fluorescence, net CO₂ assimilation, stomatal conductance, transpiration, carboxylation efficiency, and the contents of chlorophylls a and b, and carotenoids, with the strongest effects observed at 80 mM NaCl. In contrast, leaf area, biomass accumulation, specific leaf area, and intrinsic water-use efficiency were not significantly affected. Qualitative anatomical observations indicated apparent modifications in vascular organization under saline conditions, including narrower xylem vessels and phloem disorganization. Collectively, these findings demonstrate that salinity primarily impairs photosynthetic performance and vegetative growth while inducing morphological, physiological, and apparent anatomical responses in E. foetidum cultivated in vitro. This study provides the first integrated characterization of the responses of E. foetidum to in vitro salt stress, establishing a foundation for future investigations into the physiological mechanisms underlying salinity responses in this species. Full article

Anastomotic leakage (AL) and pelvic sepsis remain important causes of morbidity after restorative rectal cancer surgery, affecting mortality, stoma-free survival, bowel function, and oncologic treatment. Reported rates after low anterior resection are approximately 5–20%. We conducted a structured narrative review of PubMed/MEDLINE and the Cochrane Library through 20 July 2026, focusing on prevention, recognition, source control, and long-term recovery after low pelvic reconstruction. Most biological requirements for healing and principles of source control are approach-agnostic, whereas open, laparoscopic, and robotic surgeries create different technical conditions for pelvic exposure, stapler trajectory, articulation, tactile feedback, conversion, and fluorescence imaging. This review therefore integrates approach-specific technical constraints with post-discharge recognition, anatomy-based source control, and patient-centered long-term outcomes rather than treating these domains separately. Among preventive measures, combined mechanical bowel preparation and oral antibiotics has the strongest support in elective, non-obstructed patients; indocyanine green fluorescence is a useful adjunct when perfusion is uncertain but does not replace assessment of tension or mechanical integrity. C-reactive protein is mainly useful for ruling out major complications, whereas contrast-enhanced computed tomography remains first-line imaging and pelvic magnetic resonance imaging is best reserved for selected small defects, chronic sinuses, or complex pelvic sepsis. Management should be driven first by physiology and then by leak timing and location, defect size, conduit viability, diversion status, and cavity drainability. Durable success includes sepsis control, anatomical healing, stoma reversal when feasible, and acceptable long-term function. Full article

Background: The acromiohumeral distance (AHD) is a widely used imaging parameter for evaluating the subacromial region and shoulder function. Traditionally, it has been interpreted as a static anatomical measurement reflecting the available subacromial space and associated with rotator cuff pathology, superior humeral migration, and subacromial impingement. However, inconsistent relationships between AHD, symptoms, and function suggest that purely structural interpretations may be insufficient. Objectives: To propose a systems-based functional morphology framework that reinterprets AHD as an emergent functional state variable reflecting the instantaneous state of the glenohumeral system. Methods: A conceptual synthesis integrated evidence from shoulder biomechanics, functional morphology, movement science, connective tissue research, systems theory, and personalized neuromusculoskeletal modelling. Structural, connective tissue, neuromuscular, and mechanical determinants of AHD were integrated within a multilevel systems perspective. Results: The proposed framework conceptualizes AHD as emerging from the dynamic interaction of four organizational domains: structural constraints, connective tissue properties, neuromuscular regulation, and mechanical context. Rather than representing a fixed anatomical space, AHD is interpreted as a potential systems-level functional biomarker whose value varies according to the functional state of the glenohumeral system. This perspective explains the variability observed across imaging studies and the frequent discordance between structural findings, symptoms, and functional outcomes. Conclusions: Reframing AHD as an emergent functional state variable provides a theoretical basis for integrating anatomy, biomechanics, movement behavior, and neuromuscular control within personalized shoulder assessment. This systems-based perspective may help bridge the traditional gap between structural imaging and clinical presentation while supporting future developments in precision rehabilitation and subject-specific computational modelling. Full article

Computer-aided diagnostic technologies have demonstrated substantial advantages in 3D medical image segmentation, particularly in multimodal 3D medical image segmentation tasks, where they play a pivotal role in driving continuous innovation in related architectures. As an integration of U-Net and Transformer, the UNETR architecture has demonstrated remarkable efficacy in 3D medical image segmentation. Nevertheless, despite its successes, UNETR remains challenged by clinical complexities such as intricate tumor localization and anatomical structural diversity in complex clinical settings. To address these issues, we propose an enhanced 3D segmentation framework, UAtten-Unetr, designed to improve segmentation accuracy and robustness in complex medical scenarios. The framework captures global contextual information via hierarchical Transformer layers and incorporates a spatial–channel attention module to enable adaptive fusion of multimodal features, thereby effectively enhancing cross-modal feature alignment capabilities. Concurrently, we innovatively developed a unified loss function based on bimodal modality-specific Dice constraints and uncertainty regularization, optimized for synchronous learning across the ACDC (cardiac MRI) and AMOS22 (abdominal CT/MRI) datasets. Experimental results showed that UAtten-Unetr achieved an average Dice score of 92.20% on the ACDC dataset, exceeding the reported nnU-Net result of 91.61% by 0.59 percentage points. On the AMOS22 dataset, the proposed method achieved an average Dice score of 84.51%, exceeding the reported UNETR result of 78.33% by 6.18 percentage points. However, its myocardium Dice score (84.11%) was lower than those of nnU-Net (89.24%) and MT-UNet (89.04%), indicating a remaining limitation in myocardium boundary segmentation. These results indicate competitive segmentation performance under the reported experimental settings. This method delivers dual improvements in accuracy and generalization across complex anatomical scenarios, providing an effective solution for precise diagnosis in intricate clinical environments. Full article

Basilar trunk aneurysms (BTAs) are rare intracranial aneurysms. These aneurysms pose substantial clinical risk because they are located close to vital cranial structures and perfusion-sensitive perforating arteries. In addition, outcomes from BTA repair are dependent upon several factors, which include proximity to the brainstem, preservation of perforators, corridor access to the aneurysm, successful clipping of the aneurysm through reconstruction, the need for intraoperative rescue manoeuvres and the ability of the patient to recover from complications. Many previous studies have documented the outcomes associated with BTA repair; however, few studies have examined how the anatomy of the BTA directly relates to surgical decisions made by surgeons. Therefore, we developed an explainable AI framework for documenting surgeon reasoning regarding the open microsurgical treatment of BTAs. The primary objective was to develop and internally evaluate an explainable microsurgical intelligence framework for structuring surgeon reasoning during open microsurgical treatment of BTAs. The secondary objectives were to explore the relationships between the proposed constructs and postoperative pontine infarction, angiographic occlusion, functional outcome, hidden disability, operative difficulty, and composite technical-safety failure. Methods: We retrospectively analysed the cases of 31 adult patients who underwent open microsurgical treatment of a basilar trunk aneurysm at our hospital between October 1999 and March 2025. The cohort included 18 women (58.1%) and 13 men (41.9%), with a median age of 55 years (interquartile range, 44–63 years); 14 patients (45.2%) presented with ruptured aneurysms. A database containing more than 300 variables collected information about each patient’s imaging studies, operative strategies employed during surgery, intraoperative events encountered during surgery, intraoperative angiographic verification, occurrence of new injuries or complications resulting from surgery, and degree of recovery in each patient. Temporally separated scores were generated to quantify perforator-aware hazard, compression burden imposed by proximity to the brainstem, constraints imposed by corridors available for clipping of the BTA, burden imposed by clip reconstruction, degree of surgical precision adjusted based on the need for rescue manoeuvres, degree of dataset/model readiness, and degree of case-level learning density. The framework was evaluated internally using methods that included leave-one-out cross-validation, Firth regression modelling, Bayesian modelling, bootstrap optimism correction, calibration assessments using Brier score, decision-curve analysis, evaluations of explainability, conformal uncertainty estimation, and retrieval of similar cases. Results: Complete occlusion of the aneurysm was successfully achieved in 27 patients (87.1%), whereas residual neck or sac remained in four patients (12.9%). New pontine infarction occurred in six patients (19.4%), including four perforator-related infarctions (12.9%). A favourable last-follow-up modified Rankin Scale (mRS) score of 0–2 was achieved in 25 patients (80.6%); however, hidden disability was noted in 11 of these 25 patients (44.0%). Mortality was 6.5% and was limited to two patients with high-grade rupture. PAH-S-pre predicted pontine infarction (OR, 1.19 for every five-point increase; leave-one-out cross-validated AUC, 0.92). The composite technical-safety failure model had an ROC AUC of 0.91, optimism-corrected AUC of 0.88, calibration slope of 0.96, Brier score of 0.10, and permutation p < 0.001. Explainability indicated that PAH-S-pre, BPCI, and CCR represented the most important features. Conformal prediction resulted in abstention from prediction in four patients (12.9%). Conclusions: Expert microsurgical thought processes involved in repairing BTAs can be systematised and recorded into time-relevant and clinically meaningful measures that preserve anatomical interpretability. Clinical use will require external validation before implementation; however, this framework may provide a clinically interpretable foundation for risk-adapted planning, verification, surveillance, education, and future decision-support research for complex neurovascular surgery. Full article

Dynamic 3D Gaussian Splatting (3DGS) enables efficient facial-avatar rendering but may suffer from cumulative geometric drift and degradation of identity-specific details when applied to long monocular sequences. To address this problem, we propose IA-3DGS, an identity-anchored dynamic Gaussian framework for long-term facial animation. The proposed framework contains three main components. First, a 3D morphable model-guided semantic initialization strategy associates Gaussian primitives with anatomically meaningful facial regions. Second, a region-weighted Jacobian rigidity regularizer suppresses non-physical shear and anisotropic stretching in quasi-rigid facial regions while preserving sufficient flexibility in expression-sensitive regions. Third, a cyclic memory correction mechanism periodically aligns the current identity representation with a frozen reference representation and applies exponential moving average smoothing to reduce correction-induced temporal discontinuities. Experiments were conducted on the NeRSemble and NHA datasets and on a self-collected Custom-5K dataset containing continuous monocular facial sequences longer than 5000 frames. IA-3DGS achieved a PSNR of 31.85 dB, an SSIM of 0.942, and an LPIPS of 0.038 on standard-length sequences. At frame 5000, the method obtained an L-LPIPS of 0.046 and a landmark mean error of 1.3 mm, compared with 0.245 and 6.8 mm, respectively, for the purely implicit dynamic 3DGS baseline. The system rendered 1920 × 1080 images at approximately 48 frames per second on a single NVIDIA RTX 4090. These results indicate that the proposed semantic, geometric, and temporal constraints improve long-sequence identity consistency under the evaluated subject-specific monocular setting. Full article

Contemporary patient-specific subperiosteal implants (SPIs) have re-emerged as digitally planned, additively manufactured solutions for oral and maxillofacial rehabilitation when conventional endosseous implants are limited by severe atrophy, anatomical constraints, or reconstructive defects. Unlike conventional implants, SPIs behave as fixation-based skeletal frameworks whose performance depends on passive fit, screw fixation, anchorage, framework architecture, material properties, manufacturing accuracy, and prosthetic load transfer. This scoping review evaluated the maturity of mechanical design and validation evidence for contemporary SPIs. Following a predefined internal protocol and PRISMA-ScR, MEDLINE/PubMed, Scopus, Web of Science, Embase, and the Cochrane Library were searched from inception to 13 June 2026. Reference-list screening and citation tracking supplemented the electronic search. Two reviewers independently screened records against predefined eligibility criteria. Data were charted using a predefined extraction form and synthesized descriptively by evidence type, engineering domain, validation stage, and translational status. No meta-analysis was undertaken because of methodological heterogeneity, and no formal risk-of-bias grading was applied. Across 65 included records, the evidence was dominated by descriptive technical studies and comparative computational analyses, whereas direct mechanical testing, fatigue assessment, manufacturing verification, and clinical correlation were limited. Finite element analysis was useful for comparing design alternatives and identifying stress concentrations, but models were heterogeneous and often insufficiently validated. Design modifications generally redistributed stress across the implant–prosthesis–bone system rather than reducing it globally. Titanium and Ti6Al4V were the most established framework materials, whereas polymeric, ceramic, and scaffold-assisted strategies remained preliminary. The principal contribution of this review is a cross-domain appraisal of progression from anatomical feasibility and comparative modeling to manufacturing verification, experimental testing, and clinical validation. An evidence map, minimum reporting checklist, and integrated validation pathway are provided to support reproducible device development. Full article

Precision livestock farming (PLF) relies on high-precision three-dimensional (3D) horse reconstruction and automatic body measurement to support refined breeding management and health surveillance. However, data collection is restricted by building environment noise and hardware layout constraints; complex equine body shapes and large individual variations induce local geometric distortions in reconstructed models, limiting field deployment. Drawing on generative 3D reconstruction, this study develops the first image-to-3D pipeline that leverages three consumer depth cameras to reconstruct high-fidelity 3D horse models, integrating reconstruction, non-rigid optimisation and automatic body measurement. The image-to-3D module accurately extracts core morphological traits such as torso outlines and limb ratios for initial reconstruction. To eliminate local geometric deformation and recover scene scale, coarse-to-fine non-rigid fitting optimisation with dynamic surface feature matching weights is proposed, strengthening alignment between reconstructed meshes and real horse anatomical structures. Comparative experiments on multiple equine datasets verify that our method surpasses existing algorithms in measurement precision and reconstruction integrity. Compared with baseline methods, the non-rigid registration reduces core body measurement errors and Chamfer Distance (CD) by over 50%, while increasing the F-Score by more than 20%. This work enables automatic horse body phenotyping and offers technical references for image-to-3D dimensional measurement of other livestock species in large-scale precise breeding. Full article

Background: Tricuspid valve surgery is indicated for severe symptomatic tricuspid valve regurgitation (TR) when performed concomitant to another cardiac operation. While percutaneous therapies for TR continue to advance, anatomic constraints may necessitate consideration of surgery in some patients. Methods: Herein, we provide a contemporary review of surgical techniques used to treat TR, which can serve as a complement to existing and emerging tricuspid valve treatment therapies. Results: We performed a literature review on established surgical tricuspid repair techniques. Conclusions: Surgical repair of the tricuspid valve is safe and feasible. While annuloplasty remains the cornerstone operative strategy, leaflet and sub-valvular techniques serve as adjuncts to help successful repair. The therapeutic paradigm for TR continues to evolve. Full article

Background/Objectives: Image-guided adaptive brachytherapy is a core component of definitive treatment for locally advanced cervical cancer (LACC). However, implantation remains challenging in bulky, asymmetric, or anatomically complex tumors, where standard applicator geometry or purely straight interstitial trajectories may be insufficient for individualized target coverage. This study evaluated the real-world implementation of a patient-specific 3D-printed template-guided adaptive brachytherapy workflow for complex LACC. Methods: We retrospectively reviewed 120 consecutive patients with FIGO 2018 stage IB3–IVA cervical cancer treated with definitive chemoradiotherapy followed by high-dose-rate image-guided brachytherapy between March 2023 and March 2025. All patients were treated using a patient-specific 3D-printed template-guided hybrid intracavitary/interstitial workflow integrating CT/MRI-based target assessment, individualized catheter trajectory planning, template fabrication, implantation verification, and adaptive treatment planning. Straight-channel or curved-channel guidance was selected according to residual tumor geometry and pelvic anatomy, with flexible plastic interstitial catheters used for curved or anatomically constrained trajectories. Procedural deliverability, dosimetry, toxicity, early clinical outcomes, and exploratory dose–outcome patterns were analyzed. Results: The median HR-CTV volume was 55.9 cm³, and the median HR-CTV D90 was 92.9 Gy EQD2. Median organ-at-risk D2cc values remained within contemporary institutional and guideline-consistent constraints. A total of 555 template-guided HDR brachytherapy fractions were delivered. The median applicator-and-catheter placement time was 4.21 min per fraction, with a median of 7.25 implanted channels. Minor and major insertion-related bleeding occurred in 10.8% and 1.7% of patients, respectively. At a median follow-up of 20.1 months, estimated 3-year overall survival, progression-free survival, local recurrence-free survival, regional recurrence-free survival, and distant metastasis-free survival were 77.9%, 76.8%, 94.3%, 98.0%, and 86.2%, respectively. Late grade ≥ 3 gastrointestinal and genitourinary toxicities occurred in 2.5% and 1.7% of patients, respectively, with no grade 4–5 events. Exploratory dose–outcome analyses suggested hypothesis-generating dose–outcome patterns, but these findings were not intended to define or validate a clinical dose threshold. Conclusions: This real-world implementation study supports the feasibility of patient-specific 3D-printed template-guided adaptive brachytherapy for complex LACC. By translating CT/MRI-based individualized trajectory planning into template-guided intracavitary/interstitial catheter placement, this workflow achieved guideline-consistent target coverage, acceptable organ-at-risk doses, efficient procedural delivery, and low severe toxicity within the available follow-up. Dose–outcome findings remain exploratory and require validation in more mature cohorts. Full article

Background/Objectives: Rectal gastrointestinal stromal tumors (GISTs) near the dentate line (DL) present significant surgical challenges due to anatomical constraints and difficulty preserving anal function. Endoscopic intermuscular dissection (EID) is a viable alternative in these cases. Methods: We retrospectively reviewed the medical records of patients who underwent EIDs for rectal GISTs near the DL between January 2015 and December 2023. The primary outcomes assessed were procedural efficacy—measured by en bloc and complete resection rates—and safety, evaluated regarding delayed bleeding, perforation, and tumor recurrence. Results: In total, 10 patients underwent rectal GIST EIDs. The mean tumor size, procedure time, and resection speed were 25.70 ± 14.09 mm, 81.30 ± 46.20 min, and 16.91 ± 9.94 mm²/min, respectively. The en bloc and complete resection rates were 100% and 70%, respectively. Compared with endoscopic submucosal dissection (ESD) for rectal epithelial lesions, there were no significant differences in procedure times (81.30 ± 46.20 min vs. 76.50 ± 32.42 min, p = 0.758) or complete resection rates (70% vs. 95%, p = 0.095). Regarding safety, delayed bleeding (10% vs. 10%, p = 1.000) and perforation (10% vs. 0%, p = 0.333) rates did not differ significantly between the two groups. During a median 36-month follow-up, one recurrence occurred in a patient with a high-risk GIST. Conclusions: EID for rectal GISTs near the DL appears to be a feasible and potentially safe treatment option in carefully selected patients. It may represent a less invasive alternative to surgery, particularly when preservation of anal function is a priority. Nevertheless, validation in larger prospective studies is required. Full article

Background: The natural motion of the human knee involves a combination of rotation and translation, resulting in a variable Instantaneous Center of Rotation (ICR) throughout the gait cycle. Traditional prosthetic knee designs often fail to reproduce this complex kinematic behavior. Objectives: This study aims to propose a customized, biomimetic knee mechanism through the synthesis of a four-bar linkage capable of approximating the physiological ICR trajectory with high precision. Methods: A Genetic Algorithm (GA) was implemented to optimize the geometric parameters of the four-bar mechanism, specifically its link lengths and inter-link angles. The optimization process is based on Freudenstein’s equation, which analytically relates the input and output angles of the linkage to the lengths of its links. The desired ICR trajectory was derived from experimental data, and the objective function minimized the Euclidean error between the generated and target trajectories. Results: The proposed method yielded customized mechanisms that closely approximate the target ICR curves, achieving an overall mean Euclidean tracking error of 1.95% (±1.68%) across diverse patient profiles, with a best-case optimization error as low as 0.355%. Furthermore, the GA demonstrated high computational efficiency, converging on optimal geometric configurations in an average execution time of just 3.98 min. Conclusions: These numerical results validate the robustness of the GA in navigating the design space while strictly adhering to kinematic constraints, Grashof’s condition, and anatomical motion limits. The integration of Freudenstein’s equation with GA-based optimization techniques enables the customized synthesis of four-bar linkages with a high capacity to reproduce the physiological kinematics of the knee. This computational approach could be highly beneficial for the design of polycentric knee prostheses, as it reduces design and manufacturing time by providing the initial parameters for the development of the four-bar mechanism, ultimately ensuring a better biomechanical fit between prosthetic and natural human movement. Full article

The anuran atlas represents one of the most structurally simplified cervical systems among tetrapods, consisting of a single vertebra that must simultaneously support the head, permit mobility, and withstand mechanical loads generated during locomotion and feeding. This unique configuration provides a compelling framework for investigating how structural constraint, allometry, and ecology interact to shape morphological evolution. Here, we examine atlas shape diversification across Hyloidea using geometric morphometrics and phylogenetic comparative methods on 421 specimens representing 133 species. We quantified shape variations in dorsal and ventral views and tested the relative contributions of body size, phylogenetic history, microhabitat, and locomotor mode to atlas morphology. Atlas shape exhibited significant phylogenetic signals and was predominantly structured by allometric scaling, with the atlas size (centroid size) explaining a significant proportion of variation in both anatomical views, particularly ventrally. Morphological differences were concentrated in atlas width and relative proportions, indicating that diversification occurs primarily through size-dependent remodeling rather than through major structural reorganization. In contrast, ecological variables explained little independent variation in atlas shape, although microhabitat modulated allometric trajectories in dorsal and ventral views. Together, these results indicate that atlas evolution in hyloid frogs is governed chiefly by size-related biomechanical demands operating within a phylogenetically structured and developmentally constrained system. More broadly, our findings suggest that extreme cervical simplification channels morphological diversification toward allometric modification, limiting the extent to which ecological specialization can drive anatomical divergence. Full article

Background/Objectives: Osteoporosis is a major contributor to vertebral compression fractures (VCFs) and other skeletal complications, yet quantitative bone mineral density (BMD) assessment using dual-energy X-ray absorptiometry (DEXA) is not routinely available in many spine surgery workflows. This study proposes an anatomically guided and optimization-refined radiomics framework for opportunistic osteoporosis assessment from routine lumbar spine magnetic resonance imaging (MRI). Methods: The proposed pipeline employs a hierarchical template-matching strategy to automatically localize the L1–L4 vertebral region, followed by an optimization-based refinement procedure that adapts vertebral regions of interest (ROIs) using intensity, texture, boundary, and geometric constraints. Anatomically consistent ROIs are subsequently used for extraction of handcrafted radiomic descriptors, including statistical, textural, gradient-based, frequency-domain, and shape-related features. The extracted features were evaluated using conventional support vector classification (SVC) and a NeurodynamicSVMRBFTanh classification framework for osteoporosis-related classification. Results: Experimental results demonstrated robust and anatomically consistent vertebral localization across heterogeneous lumbar MRI acquisitions. The NeurodynamicSVMRBFTanh framework achieved the best screening-oriented performance, yielding 85.2% classification accuracy and 100.0% sensitivity on an independent test set. In addition, exploratory BMD regression analysis demonstrated the feasibility of estimating DEXA-derived BMD directly from MRI-derived radiomic features, achieving mean absolute percentage errors of approximately 15–20% across lumbar vertebral levels. Conclusions: These findings suggest that anatomically guided vertebral radiomics extracted from routine lumbar spine MRI contain clinically meaningful information associated with osteoporosis-related bone quality changes and may provide a practical tool for automated opportunistic osteoporosis assessment in settings where DEXA measurements are unavailable. Full article

Infrared thermal imaging offers objective physiological insights for Traditional Chinese Medicine (TCM), yet automated acupoint localization struggles with low texture and extreme pose variations. To address this, we constructed a multi-pose thermal facial dataset using digitized bone-proportional measurements and TCM anatomical rules. We propose T2FAL, a two-stage cascaded regression framework explicitly decoupling macroscopic face detection from fine-grained acupoint localization. Stage 1 utilizes an improved YOLOv12m-based Thermal-Aware Face Detector—integrating ICAN_C2f, MixNeck, and TAF-IoU—to mitigate domain shifts and thermal noise. Stage 2 deploys pose-specific regressors incorporating Gated Feature-Conditioned Cascade Refinement (FCCR) and a Selective GeoDeriv module, mathematically translating anatomical rules into geometric constraints. Under the stated experimental protocol, Stage 1 processed images at 87.3 frames per second on an NVIDIA RTX 4090. For Stage 2, the final frontal- and profile-view configurations achieved mAP@50–95 values of 73.19% and 86.92%, with mean pixel errors of 1.986 and 3.109 pixels, respectively. These results support the feasibility of automated reference-coordinate localization on the datasets used. However, given the partial reliance of the annotations on image registration and geometric rules, the use of the cross-domain set for model selection, and the lack of clinical or diagnostic evaluation, further validation using independently generated expert annotations, a strictly held-out external test set, and clinically labeled data is required before clinical or diagnostic use. Full article