Aldo von Wangenheim

Objectives: Gestational diabetes mellitus (GDM) is associated with fetal hyperinsulinemia, excessive adiposity, and adverse perinatal outcomes; sonographic markers of fetal adiposity, liver enlargement, and septal thickening may allow direct assessment of fetal metabolic adaptation. This study aimed to develop an ultrasound-based fetal metabolic score (FMS) estimating the risk of fetal hyperinsulinemia. Methods: The FMS was developed from a prospective cohort of 223 pregnant women (128 with GDM, 95 controls). Sonographic markers of neonatal hyperinsulinemia were identified using umbilical cord C-peptide above the 75th percentile as the reference outcome: thigh soft tissue thickness, abdominal subcutaneous fat thickness, fetal liver length, and interventricular septal thickness. Odds ratios from logistic regression were converted into weighted integer coefficients, and continuous measurements were categorized by gestational age-adjusted percentiles into a composite score. Results: Fetuses of GDM mothers showed significantly greater abdominal adiposity, liver length, and interventricular septal thickness than controls. Thigh soft tissue thickness (OR 1.62; 95 % CI 1.25–2.09; p<0.001), abdominal subcutaneous fat thickness (OR 1.57; 95 % CI 1.23–1.99; p<0.001), fetal liver length (OR 1.06; 95 % CI 1.02–1.11; p=0.004), and interventricular septal thickness (OR 1.35; 95 % CI 1.01–1.82; p=0.040) independently predicted elevated cord blood C-peptide. The FMS ranged from 0 to 44 points, integrating weighted contributions from adiposity, visceral growth, and cardiac remodeling markers. Conclusions: The FMS is a novel ultrasound-based approach for estimating fetal hyperinsulinemia risk. By integrating multiple manifestations of the fetal metabolic response to maternal hyperglycemia, it may assess fetal metabolic health more comprehensively than conventional biometry alone.

Autonomous-driving perception is typically developed and evaluated on paved roads, despite unpaved surfaces representing a large share of road networks (e.g., 78.5% in Brazil). We present a systematic literature review of computer-vision methods for detecting unpaved road distress published from January 2020 to May 2025. Following a Population–Intervention–Outcomes framework, we searched seven databases (ACM DL, arXiv, ASCE Library, IEEE Xplore, ScienceDirect, SpringerNature, and Wiley OL) using prespecified eligibility criteria. Screening yielded 1,400 records, 39 full texts, and 11 eligible studies. Deep learning is the preferred computer vision technique today, while classical methods (e.g., morphological operations) are occasionally used as support. Semantic segmentation is the most common task, followed by object detection and classification; depth and size estimation are rare. Most systems use RGB cameras, typically vehicle-mounted (roof or dashboard), with fewer studies incorporating stereo cameras, LiDAR, or low-atitude UAV setups. Pothole detection dominates the literature, whereas rutting and washboarding are seldom addressed. Overall, the reviewed works indicate that low-cost visual pipelines can support unpaved-road distress perception, but broader progress is limited by restricted data diversity, narrow task coverage, inconsistent evaluation, and incomplete reporting of methodological details.

Autonomous-driving datasets under-represent unstructured, mixed-surface coastal roads, especially beaches and coastal roads where multiple sand types coexist. We identified three gaps in current unstructured environment datasets: (i) the absence of sand-based terrain within unstructured environments along with dirt and poorly maintained paved roads, (ii) the lack of distinction among sand terrains (soft, hard, and wet) despite their different traversability, and (iii) the scarcity of high-resolution data representing sand roads. To address these gaps, we introduce the Floripa Coast Driving Dataset (FCDD), a high-resolution (4K) semantic-segmentation dataset covering a 12.2 km route across paved, dirt, and drivable beach roads. FCDD provides 500 annotated images (400 train / 100 val) with 17 classes organized into surfaces (asphalt, dirt road, soft sand, hard sand, wet sand), signs (markings, cat’s eyes, retaining walls, speed bump), distress (cracks, potholes), and things (ego, vehicles, people, animals, obstacles). We also study the role of image resolution using four segmentation models (DeepLabV3+, HRNet, Mask2Former, PP-LiteSeg) across 256–2048 px. Performance improves markedly from 256 px to 512 px and continues up to 1536 px; small-sized classes (e.g., markings, cat’s eyes, cracks) benefit the most from higher resolution, whereas surface classes show limited gains. These results indicate that ultra-high resolutions are unnecessary for differentiating road surfaces (512 px suffices), while mid-to-high resolutions (1536 px) maximize accuracy on small-sized categories.

The development of new technologies, improved by advances in artificial intelligence, has enabled the creation of a new generation of applications in different scenarios. In medical systems, adopting AI-driven solutions has brought new possibilities, but their effective impacts still need further investigation. In this context, a chatbot prototype trained with large language models (LLMs) was developed using data from the Santa Catarina Telemedicine and Telehealth System (STT) Dermatology module. The system adapts Llama 3 8B via supervised Fine-tuning with QLoRA on a proprietary, domain-specific dataset (33 input-output pairs). Although it achieved 100% Fluency and 89.74% Coherence, Factual Correctness remained low (43.59%), highlighting the limitations of training LLMs on small datasets. In addition to G-Eval metrics, we conducted expert human validation, encompassing both quantitative and qualitative aspects. This low factual score indicates that a retrieval-augmented generation (RAG) mechanism is essential for robust information retrieval, which we outline as a primary direction for future work. This approach enabled a more in-depth analysis of a real-world telemedicine environment, highlighting both the practical challenges and the benefits of implementing LLMs in complex systems, such as those used in telemedicine.

Objective: To investigate the association between prenatal ultrasonographic markers of macrossomia and C-peptide, a neonatal hyperinsulinemia marker, in pregnancies complicated by gestational diabetes mellitus (GDM), with a focus on fetal adipose tissue thickness, liver length, and interventricular septal thickness. Methods: This prospective cohort study included 223 pregnant women followed from 28 to 36 weeks of gestation in two referral centers in Brazil. The GDM group and matched controls underwent serial ultrasound assessments of fetal biometry, including thigh, abdominal, and subscapular skinfolds, fetal liver length, and interventricular septum thickness. Neonatal hyperinsulinemia was assessed using umbilical cord C-peptide levels. Statistical analyses included t-tests, chi-square tests, correlation analyses, and multivariate logistic regression. Results: Fetuses of mothers with GDM exhibited significantly greater abdominal [t(221) = −3.999, p < 0.01] and subscapular [t(221) = −2.502, p = 0.02] skinfolds, liver length [t(221) = −3.785, p < 0.01], and interventricular septum [t(221) = −4.781, p < 0.01] thickness. However, umbilical cord C-peptide levels did not differ significantly between groups [t(189) = −1.724, p = 0.09]. Only weak correlations were found between fetal ultrasound markers and C-peptide levels. Among all parameters, subcutaneous tissue thickness showed the highest (ρ = 0.30), though still limited, predictive value. Conclusions: Fetuses of mothers with GDM demonstrated increased measures of liver length, subscapular adiposity, and interventricular septal thickness compared to controls. However, these prenatal biometric markers showed weak correlations with neonatal C-peptide levels.

While routine macroscopic placental assessment is performed in all deliveries, anatomopathological examination is reserved for selected cases due to logistical and resource constraints. Although artificial intelligence (AI) has shown promise in medical imaging, its application to postpartum placental screening remains underexplored in clinical settings. This study addresses this gap by developing a deep learning pipeline to improve segmentation of placental structures using macroscopic photographs captured in delivery rooms, with the aim of supporting anatomopathological referral decisions. A proprietary dataset was constructed under both standardized and routine clinical conditions, with annotations curated by pathologists across eight morphological classes. Three architectures—YOLOv11, U-Net with ResNet34, and U-Net with EfficientNet-B0—were evaluated for segmentation and classification performance. Results demonstrated moderate efficacy: YOLOv11 achieved an AP50 of 72% for instance segmentation, while U-Net variants showed higher classification accuracy (up to 85.18%) but lower segmentation precision (Dice scores: 24.70–61.97%). These findings support the feasibility of incorporating AI-based photographic analysis into routine clinical workflows for placental evaluation. Despite preliminary segmentation limitations, this work establishes a foundation for AI-driven placental screening, offering potential to optimize resource allocation by guiding anatomopathological referrals.