Search Results (1,863)

Unmanned Aerial Vehicles (UAVs) have become critical platforms for missions such as surveillance, reconnaissance, and target tracking, which require real-time decision-making, reliable sensing, and efficient resource utilization. However, limited onboard computing capacity, energy constraints, variable terrain conditions, and situations where targets are partially or fully obscured limit the performance of traditional fixed-configuration sensing and tracking approaches. In this study, we propose a context-aware and adaptive UAV-based target detection and tracking framework that dynamically selects the most appropriate detection and tracking algorithm by jointly evaluating terrain characteristics and mission requirements. The proposed system includes a three-stage terrain analysis module supported by HSV color space filtering, Canny edge detection, Laplacian texture variance, and contrast-based features. In cases where color-based classification is insufficient, Random Forest-based classification is used to distinguish between vegetation, bare ground, and urban areas; the terrain classification model achieves approximately 90% accuracy during the training and testing process. In the target detection phase, a YOLOv11-based model was trained on a specialized tank dataset created from various sources and labeled in YOLO format, achieving an mAP50 performance of approximately 85%. In the tracking phase, single-object and multi-object tracking algorithms are selected via a scoring-based decision mechanism depending on the terrain type and mission scenario. Additionally, a hybrid anomaly detection mechanism that evaluates target loss, sudden bounding box changes, and view inconsistencies was integrated into the system, thereby enhancing tracking reliability and enabling the re-detection or algorithm switching process when necessary. Experimental results demonstrate that the proposed context-aware approach can reduce computational load while maintaining tracking robustness under various environmental conditions. These findings highlight that environmental awareness and adaptive algorithm selection can make significant contributions to autonomy, operational efficiency, and real-time reliability in UAV-based imaging systems. Full article

Monitoring the vertical progress of reinforced-concrete buildings supports construction management, urban analytics, and seismic exposure classification, yet camera-based floor counting faces two obstacles: public datasets depict almost exclusively completed structures, and the number of structurally finished floors is visually ambiguous while a building is still being erected. We reformulate building-height estimation as discrete floor-count classification from a single street-level facade image and assemble a 29,049-image multi-source corpus centered on the reinforced-concrete urban stock of Kazakhstan, including a 12-month, fixed-viewpoint sequence of 2255 frames that isolates invariance to construction stage, illumination, weather, and season. We formalize a reproducible annotation protocol for three recurring structural ambiguities—incomplete upper floors, rooftop superstructures, and open ground-level pilotis—and propose DINOv2-MSTS, a dual-branch architecture that aggregates multi-scale patch-token statistics from a frozen self-supervised backbone, trained with an Ordinal-Aware Annotation-Uncertainty (OAU) loss for which its Gaussian spread is learned rather than fixed. On the 5359-image Korter + Mendeley 21-category benchmark, the model attains 80% top-1 accuracy, 94% within $\pm 1$ floor accuracy, and 0.28-floor mean absolute error on this saturated 21-category task (a lower bound for buildings of 21 or more floors) using only 1.84 M trainable parameters, 165× fewer than a fully fine-tuned Vision Transformer, which it outperforms by eight accuracy points. On the separate 2255-frame IITU fixed-label robustness probe, it preserves the correct six-floor prediction in 91% of frames (0.09-floor MAE). The corpus, protocol, architecture, and loss together provide a reproducible benchmark for construction-stage building monitoring. Full article

Objectives: This retrospective study investigated the incidence and risk factors of dry socket following mandibular third molar extraction. Methods: This retrospective analysis utilized clinical information and diagnostic X-ray images of patients who underwent tooth extraction by a single oral and maxillofacial surgeon between June 2020 and March 2024. Dry socket was diagnosed according to established clinical criteria. Panoramic radiographs were used to assess angulation (Winter’s classification), depth and gingival coverage. Statistical analyses were conducted to examine differences in dry socket incidence. Categorical comparisons were performed using chi-square or Fisher’s exact tests. Multivariate logistic regression was used to identify independent risk factors for dry socket. Odds ratios with 95% confidence intervals were calculated. Results: A total of 3033 mandibular third molars were extracted from 2384 patients (859 males and 1525 females). The overall incidence of dry socket was 2.6% (80 cases), with a significantly higher rate in female patients (3.3%) than in male patients (1.4%) (p < 0.01). Increased risk was observed in age ≥ 30 years compared to early 20 s (p < 0.01), and the use of oral contraceptives (p < 0.01). Multivariate logistic regression also identified these three factors as significant independent risk factors. Conclusions: These findings suggest that patient-related factors influencing healing capacity contribute to dry socket development. Recognizing these associations may assist clinicians in identifying high-risk cases and implementing preventive strategies to reduce postoperative complications in third molar surgery. Full article

Background/Objective: Ultrasound is the first-line imaging modality for thyroid nodule assessment; however, it remains highly operator-dependent and subject to interobserver variability. Artificial intelligence (AI)-based systems have been proposed to improve reproducibility, yet evidence regarding their agreement with clinician assessment—particularly at the level of individual sonographic features—remains limited. Importantly, most available studies evaluate concordance rather than true diagnostic accuracy against an independent reference standard. To evaluate agreement between an AI-based ultrasound system and expert clinician assessment in thyroid nodule evaluation, focusing on concordance of size measurements, agreement in sonographic feature classification, and exploratory diagnostic performance relative to cytological outcomes. Methods: This retrospective single-center study included 74 thyroid nodules from adult patients undergoing routine ultrasound examination. Archived ultrasound images were independently assessed by an experienced clinician and an AI-based system. Agreement for quantitative measurements was evaluated using Bland–Altman analysis, while categorical features were assessed using percent agreement and Cohen’s kappa coefficients. Calibration was examined using scatter plots with the line of identity. Cytological results, when available, were used as a non-uniform exploratory reference standard for diagnostic analyses. Exploratory diagnostic performance was assessed using receiver operating characteristic (ROC) curves and area under the curve (AUROC) estimates. Given the study design, analyses primarily reflect agreement and measurement concordance rather than true diagnostic accuracy. Results: AI-derived and clinician measurements demonstrated strong agreement across all dimensions, with minimal systematic bias and stable calibration patterns. A small but consistent underestimation of one measurement axis by approximately 1 mm was observed. For categorical features, agreement ranged from fair to moderate (κ = 0.196–0.368), with the highest concordance for echogenic foci and lowest for echogenicity. Exploratory analyses showed variable diagnostic discrimination, with the best performance observed for size measurements and selected sonographic features. Conclusions: AI-based ultrasound analysis demonstrates robust agreement with clinician assessment for quantitative thyroid nodule measurements, while agreement for categorical feature classification remains moderate and variable. The findings highlight that the present study evaluates concordance rather than definitive diagnostic accuracy, particularly given the lack of a uniform independent reference standard. These results support the role of AI as an assistive tool in thyroid ultrasound practice, improving measurement reproducibility while requiring ongoing clinician oversight for qualitative interpretation. Full article

Background/Objectives: Accurate estimation of meal composition from food images can support safer and more reliable insulin bolus decision-making for individuals with Type 1 diabetes. Existing food recognition and nutrition estimation systems are often designed for general dietary logging and do not directly integrate food analysis with personalized insulin therapy parameters. Methods: This study presents an image-based nutrition estimation and insulin decision-support module developed within the AI-assisted Diabetes Care (AIDCARE) platform. The proposed system uses a convolutional neural network (CNN) to classify food items from a single meal image, and retrieves reference nutritional values from a food composition database. A separate multimodal large language model (MLLM)-based estimation component is then used to estimate portion size, allowing carbohydrate and nutrient values to be scaled according to the observed serving. Results: A curated food image dataset containing 40 food categories was used to evaluate three CNN architectures: ResNet50, Inception V3, and EfficientNet-B0. EfficientNet-B0 achieved the best classification performance, with 94.91% validation accuracy, 95.55% precision, 94.87% recall, and 94.90% F1-score. The portion-estimation component achieved an MAE of 12.27 g and an RMSE of 15.11 g. The estimated carbohydrate value is combined with user-specific clinical parameters, including the insulin-to-carbohydrate ratio and insulin sensitivity factor, to generate advisory bolus guidance. To support safety, the system requires user confirmation or correction of the recognized food category and estimated portion before insulin guidance is displayed. Conclusions: The proposed system is intended for advisory decision support only and is not designed to replace clinical judgment or autonomous insulin delivery systems. Full article

Background: This study aimed to determine whether incorporating radiologic contrast during the excretory (urographic) phase enhances the detection of recurrence on [68Ga]Ga-PSMA-11 PET/CT in patients with biochemical relapse (BCR) following radical prostatectomy (RP). Methods: A single-center retrospective analysis included 43 men with BCR after RP who underwent [68Ga]Ga-PSMA-11 PET/CT. Each patient underwent two comparative assessments. In the first assessment, whole-body PET images acquired at 60 min post-injection were fused with the non-contrast CT from early dynamic pelvic imaging (PET/CTd), and local recurrence and pelvic nodal involvement were evaluated according to PROMISE V2 and E-PSMA frameworks by two blinded readers. In the second assessment, the same PET dataset was fused with the excretory-phase CT urography (CT-U) obtained during the same imaging session at 60 min post-injection, and the same parameters were re-evaluated. Endpoints included surgical-bed classification, peri-ureteric nodal status, reader confidence, ureter visualization/opacification, and interpretation time. Inter- and intra-observer agreement was assessed, and discrepancies were resolved by consensus. Results: Surgical-bed positivity decreased from 12/43 (27.9%) on PET/CTd to 5/43 (11.6%) on PET/CT-U, leading to reclassification in seven patients (p = 0.016). Reader confidence improved significantly in five cases (p < 0.005). Peri-ureteric nodal status was changed in four patients (two positive-to-negative and two negative-to-positive), with overall positivity unchanged (5/43 vs. 5/43; p = 1.000). Ureter visualization improved markedly (inadequate: 31 vs. 10 cases), reducing diagnostic uncertainty by 50%. CT-U opacification was ≥50% in most cases (κ = 0.814), enabling reliable delineation of the ureteral course. Inter-reader agreement remained strong (surgical bed κ: 0.944 vs. 0.876; nodes κ: 0.896 both). Interpretation time decreased for both readers (senior: 3.12 vs. 2.10 min (−32.7%); junior: 4.06 vs. 2.42 min (−40.4%)). Conclusions: Adding an excretory-phase CT urography to [68Ga]Ga-PSMA-11 PET/CT improves diagnostic confidence, reduces interpretive uncertainty in the surgical bed, clarifies peri-ureteric nodal findings, enhances ureter visualization, and shortens interpretation time. CT-U is a practical enhancement to low-dose PET/CT protocols for BCR after RP. Full article

The BQ method, the most widely used technique for classifying tunnel surrounding rocks, requires correction factors including groundwater conditions, structural surface characteristics, and initial ground stress. However, existing deep learning approaches address only a single parameter and rely heavily on empirical judgment, lacking the capability for comprehensive multi-parameter intelligent assessment. To address these limitations, a large-scale database of 21,800 tunnel face images was constructed through on-site data collection, data aggregation, and image augmentation. Five convolutional neural network (CNN) models were trained and evaluated using a proposed multi-indicator scoring method comprising seven performance metrics: loss value, accuracy, precision, recall, confusion matrix, frames per second (FPS), and model size. EfficientNet-B2, ResNet101, and DenseNet121 achieved the highest scores of 100, 84, and 87 for groundwater classification, rock structure type classification, and weathering degree classification, respectively. These three optimized models were integrated into a unified software platform that recognizes multiple geological attributes from a single tunnel face image. Field validation across multiple sections of the Jiaozhou Bay Second Undersea Tunnel shows that the platform achieves a recognition accuracy of over 90%. The results demonstrate that the proposed multi-indicator evaluation method yields more-comprehensive model selection, and the integrated platform can directly support BQ value correction, contributing to intelligent surrounding rock classification in complex tunnel construction environments. Full article

Accurate hyperspectral image classification is fundamental to geospatial applications but is often constrained by annotation scarcity. To achieve high classification performance under small-sample conditions, we propose the Memory-Guided Adaptive Spectral–Spatial Perception model, which incorporates a three-level globalization strategy. At the single-sample level, an adaptive perception Transformer combines deformable and dilated convolutions with a Transformer encoder to capture global context from individual samples. At the intra-batch level, we introduce a metric learning strategy that explicitly captures structural dependencies and feature relationships among samples within each mini-batch, enabling comprehensive feature aggregation in a localized context. At the cross-batch level, a memory-guided strategy constructs a dynamic memory bank to store and retrieve features from same-class samples across training iterations, bridging past and present distributions to enhance generalization. Using only 1% of the SaliLMSS, Pavia University and Kennedy Space Center datasets and 0.5% of the WHU-LongKou dataset as training samples, our method achieves outstanding overall accuracy of 96.15%, 97.81%, 89.22% and 99.32%, respectively, outperforming existing methods. Full article

This study proposes a deep learning-based framework for the automated classification of photovoltaic solar cells as defective or normal using electroluminescence (EL) imaging. A balanced dataset containing 20,400 EL images, comprising 10,200 defective and 10,200 normal solar cells, was used for model development and evaluation. To reflect practical inspection requirements, cracked and broken cells were combined into a single defective category, resulting in a binary classification task. The dataset includes both monocrystalline and polycrystalline solar cells, which were analyzed together within a unified classification framework to improve applicability to real-world photovoltaic systems. To ensure a fair and unbiased evaluation, dataset partitioning was performed prior to any preprocessing or augmentation operations, and each image was assigned exclusively to the training, validation, or test subset. Data augmentation was applied only to the training set, eliminating the possibility of data leakage. Four state-of-the-art deep learning architectures, EfficientNet-B2, ConvNeXt-Tiny, MaxViT-T, and ResNet-50, were trained and evaluated under identical experimental conditions using the same preprocessing pipeline, training strategy, and dataset split. Model performance was assessed using accuracy, precision, recall, F1-score, ROC-AUC, confusion matrices, and explainability-based activation and attention heat maps. All evaluated models achieved classification accuracies exceeding 98%, demonstrating strong capability for EL-based defect detection. EfficientNet-B2 achieved the highest numerical performance, reaching 99.31% accuracy, 0.9931 F1-score, and 0.9987 ROC-AUC. MaxViT-T exhibited similarly strong performance with rapid convergence and balanced class-wise metrics, while ConvNeXt-Tiny and ResNet-50 also produced highly reliable results. Heat map visualizations revealed that EfficientNet-B2 and MaxViT-T concentrated their attention more precisely on defect regions such as cracks and fractures, providing visual interpretability in addition to quantitative performance. The results demonstrate that modern deep learning architectures can accurately and reliably detect photovoltaic cell defects from EL images under a unified binary classification framework. Furthermore, explainability techniques enhance the transparency of model predictions, supporting the practical deployment of intelligent inspection systems for photovoltaic manufacturing and maintenance applications. Full article

The automated classification of physical violence in images faces a critical methodological obstacle: class imbalance, which undermines the discriminative capacity of Deep Learning (DL) models by leading them to overfit the majority class. Although numerous balancing strategies have been proposed in the literature, which range from traditional resampling to advanced generative models, a systematic and controlled empirical evaluation of their behavior in the specific domain of physical violence detection remains absent. In this study, we present a systematic empirical review that evaluates traditional balancing techniques, generative models, and data augmentation methods under controlled imbalance conditions. The methodology is applied to five datasets (UBI-FIGHTS, RLVSD, RWF-2000, RLVS, and AIRTLAB), across six imbalance scenarios defined by Balance Percentage levels of 1%, 5%, 10%, 25%, 50%, and 75%. To isolate the effect of the balancing strategies from architectural confounders, a single stable backbone (ResNet-18) is employed as a controlled classifier across all experimental conditions. The experimental results demonstrate the absence of a universally optimal balancing technique: in highly imbalanced scenarios, traditional oversampling methods such as ROS achieved the best average performance, while generative approaches such as DCGAN and DeepSMOTE became increasingly competitive as class balance is improved. These findings confirm that the effectiveness of balancing techniques depends on both the degree of asymmetry and the dataset’s intrinsic characteristics. Thus, it is clear that there is a need for context-aware strategy selection rather than one-size-fits-all solutions. Beyond the empirical findings, this work provides a structured synthesis of the theoretical foundations and state-of-the-art methods for imbalanced violence detection. Full article

Artificial intelligence-assisted food detection is increasingly moving from single-task classification toward integrated analytical systems capable of producing multiple quality-related outputs from one sensing workflow. However, most hyperspectral food detection studies still rely on a single spectral range or simple feature concatenation, which limits their ability to exploit complementary physicochemical information from heterogeneous sensors. In this study, an artificial intelligence-enabled visible–near-infrared and short-wave infrared (VNIR-SWIR) hyperspectral fusion framework is proposed for multi-task fruit detection, using origin authentication and bruise localization as representative tasks. The proposed method first constructs an observation-consistent fused representation from high-resolution VNIR images and low-resolution SWIR images. Collaborative spectral unmixing is used to couple cross-modal material distributions, while abundance-consistency and downsampled observation-consistency constraints are introduced to estimate SWIR-informed features on the VNIR spatial grid without assuming measured high-resolution SWIR ground truth. The fused representation is then processed by a shared spectral–spatial deep encoder with two task-specific heads: a fruit-level classification head for origin authentication and a pixel-level segmentation head for bruise detection. Experiments on apple and kiwifruit datasets show that the proposed framework outperforms VNIR-only, SWIR-only, bicubic-fusion, CNMF-style fusion, and TV-regularized fusion baselines under five fruit-level stratified random splits. For origin-category authentication, the proposed method achieved an accuracy of almost 93.85 for apples and almost 94.35 for kiwifruit. For bruise localization, the proposed method achieved higher overall accuracy, average accuracy, and Cohen’s kappa across the evaluated fruit categories. Full article

Oscillatory neural networks (ONNs) provide a neuromorphic computing framework that exploits the phase dynamics of coupled oscillators for parallel and energy-efficient pattern recognition. In this study, we design a single-layer, fully connected ONN to classify handwritten digits from the MNIST dataset. Input images were downsampled to 6 × 6 binary patterns, which were optimized using a genetic algorithm to evolve effective templates, as experiments with higher-resolution inputs showed only marginal accuracy improvements at significantly increased computational and energy costs. Coupling weights were determined using Hebbian learning, and the network dynamics were simulated using the Kuramoto model to encode information via phase relationships. To the best of our knowledge, this is the first work to apply genetic algorithm optimization to design the templates used by an ONN and to combine evolutionary template generation with Hebbian-based ONN training for image classification. The results show that the ONN achieves 75–76% accuracy in the full 10-class MNIST task, with outputs exhibiting stable sinusoidal behavior and resilience to moderate noise. These findings highlight the potential of ONNs as a practical, low-power alternative to conventional deep learning models, particularly for real-time edge-level applications where energy efficiency and robustness are critical. Full article

Purpose: To determine the prevalence and subtype distribution of accessory navicular (AN) on adult ankle MRI and to evaluate MRI factors associated with AN-related bone marrow edema (BME); bilaterality was explored in patients with available contralateral MRI data. Methods: In this retrospective cross-sectional study, consecutive adult patients (≥18 years) who underwent ankle MRI between January 2022 and January 2025 at a single institution in Van, eastern Türkiye, were identified from the institutional archive. Two blinded readers assessed AN presence, classified subtypes using Coughlin criteria, measured maximal ossicle dimension for types 1–2, and recorded BME on fluid-sensitive sequences. Prevalence with 95% CIs was estimated, and multivariable logistic regression assessed factors associated with BME in patients with types 1–2. Results: Among 1988 unique patients (mean age: 42.2 ± 14.2 years; 42.6% male and 57.4% female), AN prevalence was 24.4% (95% CI: 22.6–26.4%), and type 2 predominated (50.2%). In the subgroup of AN-positive patients with available contralateral MRI data, AN was bilateral in 61 of 69 patients (88.4%). Interobserver agreement was excellent for AN detection and subtype classification (κ = 1.00 and κ = 0.98). Among types 1–2 included in the BME analysis (n = 385), BME occurred in 81 patients (21.0%). In the multivariable model, the adjusted ORs for BME were 4.46 for type 2 morphology, 1.24 per 1 mm increase in maximal dimension, 2.08 for female sex, and 0.71 per 10-year increase in age; concomitant os trigonum was not associated with BME. Conclusions: AN was common on ankle MRI, and type 2 was most strongly associated with BME; however, because this was a retrospective, cross-sectional, symptom-blinded study, causality and clinical correlation could not be established. Clinical Significance: Type 2 AN, particularly with a larger ossicle dimension, should prompt careful evaluation for MRI-detected BME, which should be interpreted as an imaging finding requiring clinical correlation. Full article

Background/Objectives: Automated cephalometric landmark detection using deep learning has the potential to streamline routine orthodontic diagnosis. However, the clinical relevance of artificial intelligence (AI) localisation accuracy depends on how detection errors propagate into derived angular measurements and skeletal classifications. We retrospectively evaluated 14 YOLO-based model configurations and quantified the agreement between AI-derived and expert-derived ANB-based skeletal classifications. Methods: Twelve working YOLO-based models (YOLOv5xu, YOLOv11 nano/small/medium/large variants) were trained on a single-centre dataset of 120 lateral cephalograms and evaluated on an independent test set of 11 cephalograms (stratified across skeletal Classes I, II, III). The four ANB-defining landmarks (Sella, Nasion, A-point, B-point) were the focus of the analysis. Each test cephalogram had been annotated by four orthodontists (44 measurements per image), yielding the expert reference. We assessed the effects of architecture, bounding-box size (40/100/150 px), training dataset scale (235–4255 images) and training epochs on localisation accuracy (mean radial error, MRE; Successful Detection Rate, SDR) and on the downstream ANB-based skeletal classification. Diagnostic concordance was quantified by classification agreement, Cohen’s $\kappa$ with bootstrap 95% confidence intervals (10,000 iterations), an exact one-sided binomial test for discordance, and Wilson exact CIs per class. Results: The best-performing model (Model 2; YOLOv11l, 40 × 40 px bounding box, 1175 training images) achieved an MRE of $3.10\pm 1.00$ mm and a SDR@4 mm of 87.2% for S, N, A, and B. ANB-based skeletal classification demonstrated 96.9% concordance with expert assessments (95% bootstrap CI: 93.8–99.2%; Cohen’s $\kappa$ = 0.946 [95% CI 0.89–0.99]; exact binomial test against a 90% concordance threshold $p=0.003$). Per-class concordance was Class I 95.8% (23/24), Class II 94.9% (56/59), and Class III 100% (47/47). Three of four discordant cases clustered near the Class I/II diagnostic threshold (expert ANB $\approx 4.5°$). Bounding-box size dominated localisation accuracy, with a 3.5-fold increase in MRE from 40 × 40 to 150 × 150 px configurations and SDR@4 mm collapsing from 82.8% to 0%. Conclusions: Within the constraints of a retrospective single-centre design with a small (n = 11) independent test set, YOLO-based AI landmark detection demonstrated promising diagnostic concordance with expert consensus for ANB-based skeletal classification. These findings warrant prospective, multi-centre external validation before clinical deployment and support a confidence-aware workflow in which AI predictions for borderline ANB values undergo mandatory clinician verification. Bounding-box calibration emerged as the single most impactful preprocessing decision. Full article

Phosphene-based visual neuroprostheses require encoding schemes that preserve task-relevant information while remaining feasible under safety-constrained stimulation. This study proposes a stress-based evaluation framework that reframes phosphene encoder assessment as a tri-objective operating-envelope problem rather than a single-metric comparison. Four representative encoders—rate, sparse, temporal, and optim—were evaluated under structured perturbations using two simulated prosthetic-vision(SPV) benchmarks: EMNIST Letters for symbolic recognition and a COCO-derived balanced subset for a reduced four-class COCO-derived image-level classification task. The final experimental configuration used 20,000/3000/3000 train/validation/test samples for EMNIST and 4000/1000/1000 for the COCO-derived benchmark. The framework combined structured stress sweeps, residual proxy safety-burden analysis, topology-based representation metrics, and stress-integrated utility summaries within a simulation-based evaluation setting. The results showed that clean-conditioning accuracy alone was insufficient to predict encoder behavior under stress. EMNIST exhibited broad operator-dependent degradation, whereas the COCO-derived benchmark showed a lower but more compressed performance regime. Residual proxy safety burden was only loosely aligned with performance, with moderate dissociation between performance and residual proxy safety burden in EMNIST and weaker alignment between these two axes in the COCO-derived benchmark. In the point-estimate utility summaries, the sparse encoder tended to yield comparatively favorable tri-objective utility values within the present single-run simulation-based framework, simplified SPV percept-synthesis operator, and fixed benchmark-specific decoder setting, primarily because it maintained an almost negligible residual proxy safety burden while preserving competitive performance and topology-based representation metrics. Topology-based analysis further indicated that topology-based representation metrics largely tracked task degradation in EMNIST, whereas topology-based representation metrics showed larger relative variation than decoder accuracy within the evaluated simulation setting under degraded COCO-derived conditions. Taken together, these findings provide an exploratory, benchmark-specific assessment suggesting that phosphene encoder evaluation may benefit from a multi-axis operating-envelope-oriented analysis that jointly considers stressed functional performance, residual proxy safety burden, and topology-based representation metrics within the present simplified SPV and fixed-decoder evaluation setting. These results should therefore be interpreted as simulation-level, configuration-dependent observations under a simplified SPV percept-synthesis operator, with safety-related quantities treated as residual proxy safety-burden summaries rather than as direct physiological, electrochemical, clinical, or implant-specific safety measurements. Full article