Detecting the oxidative lesion 8-oxo-7,8-dihydro-2’-deoxyguanosine (8-oxo-dG) directly from nanopore signals is difficult because lesions are rare, the ionic current is noisy, and the signal perturbation depends on sequence context. We present OxoNet, a dual-branch network that separately encodes the raw current and the basecalled sequence and fuses them through cross-attention. To improve transfer beyond synthetic oligonucleotides, OxoNet is trained with synthetic positives and formamidopyrimidine-DNA glycosylase (Fpg)-repaired human genomic negatives under a grouped, read-disjoint split, following the design of the source study. Because default probability thresholds are unsuitable at lesion prevalences of 10−6–10−4, all primary comparisons are made at matched false-positive rates on held-out genomic negatives, using thresholds fixed on validation data. At a genomic false-positive rate of 10−4, OxoNet achieves 55.9% recall, against 32.0% for a gradient-boosted-tree control, 28.8% for esox and 28.1% for NanoCon-7 on identical reads and positions, and attains higher recall in all 100 sequence contexts examined. On 176,567,845 interrogated guanines of native human DNA, thresholds set to a common validation-derived recall yield 27.9 calls per million for OxoNet, 1,125.2 for NanoCon-7 and 2,286.1 for esox, and OxoNet reproduces the guanine-cytosine (GC)-dependent call-rate pattern reported in the source study. Native data carry no site-level truth and come from the same public project that supplied the synthetic training data; these results therefore characterize transfer and distribution shift, not higher native specificity or independent validation.
This work investigates the impact of Schottky barrier diode (SBD) integration on the reliability of 1.2 kV SiC MOSFETs under high-current surge stress. Due to its structural uniqueness, the SBD embedded MOSFET exhibits a distinct reverse conduction behavior. In this structure, the integrated SBD dominates the current flow and effectively suppresses the channel current compared to conventional devices. We conducted surge current experiments to analyze how this specific conduction characteristic influences device reliability. The VSD-ISURGE hysteresis loops were analyzed to evaluate energy dissipation in SBD embedded SiC MOSFETs across various surge magnitudes and compared with those of conventional devices. Additionally, mixed-mode simulations using TCAD were conducted to analyze the current density and temperature distributions under high-current surge stress. The integrated SBD facilitates current redistribution and reduces the source-drain voltage drop. This conduction mechanism results in lower energy dissipation compared to conventional devices. The SBD embedded MOSFET exhibits approximately 1 J less energy dissipation and a minimal threshold voltage shift of only 0.025 V compared to the conventional device. These results demonstrate that SBD integration significantly enhances device reliability under high-current surge stress.
Accurate polyp segmentation in colonoscopy images is critical for early detection and prevention of colorectal cancer, yet remains challenging due to variable polyp appearance, size, and boundary ambiguity. While recent transformer-based encoders such as Mix Vision Transformer (MiT) have demonstrated strong feature extraction capability, they are typically paired with decoder architectures originally designed for CNN feature hierarchies, creating a potential feature utilization bottleneck. Through systematic comparison of seven encoder–decoder combinations, we find that while MiT-B2 improves Dice by 3.8–20.7% over ResNet50, the decoder choice introduces up to 4.1% per-dataset Dice variation (0.7% on average) with the same encoder, revealing decoder design as an underexplored bottleneck. Based on this analysis, we propose a Progressive Hierarchical Decoder (PHD) with progressive channel projection and efficient skip fusion tailored for transformer feature hierarchies. PHDSeg achieves 0.918 mean Dice on Kvasir-SEG and 0.903 on CVC-ClinicDB over three independent runs, outperforming the best baseline trained under identical conditions by 1.3% average Dice across five benchmarks. We additionally investigate frequency-domain feature enhancement and discover that high-frequency augmentation improves challenging datasets (ETIS: +1.7%) but degrades well-defined cases (CVC-ClinicDB: –6.3%), providing practical guidance for frequency-based methods in medical image segmentation.
This paper presents an automatic license plate recognition framework for Azerbaijani plates featuring annotation-free character recognition, addressing limited supervision, severe class imbalance, and unseen character combinations in low-resource settings. Unlike conventional approaches that require extensive character-level annotations, the proposed method uses synthetically generated images from generative adversarial networks to support few-shot and zero-shot learning without manual labeling. The study evaluates detection, generation, and recognition jointly. Detection compares architectures under public pre-training, lightweight fine-tuning, and training from scratch on approximately 370,000 public and 412,000 private images. Generation compares four network types under controlled scaling; recognition measures synthetic-to-real transfer across fourteen dataset configurations. Five epochs of fine-tuning match or exceed training from scratch; YOLO11-S reaches mean average precision of 0.9637 averaged over intersection-over-union thresholds from 0.5 to 0.95. Image quality metrics poorly predict recognition utility: the Wasserstein generative adversarial network with gradient penalty scores best on perceptual metrics yet yields the weakest recognition. Generator diversity dominates, with multi-generator ensembles reaching average precision of 0.714 at intersection-over-union 0.5 without real character annotations. These findings support an efficient annotation-free pipeline in low-resource settings. Source code for detection, synthetic generation, and recognition is publicly available; the repositories do not redistribute the datasets used in this study.
This paper proposes a system for situation understanding in multi-LiDAR monitoring that transforms person-level observations into a structured spatio-temporal and relational representation for reasoning with a large language model (LLM). The representation organizes observations within temporal windows as individual events and interpersonal relations. We evaluate the proposed system on individual action classification and pairwise social group classification using person observations derived from JRDB annotations. The results show that supervised classifiers using compact motion features are more effective for individual action classification. For social group classification, an LLM using structured event and relation descriptions achieves higher Macro-F1 and minority-class F1 than one using compact numerical summaries, while showing no clear paired differences from the supervised baselines despite requiring no task-specific training. The method also yields the best group reconstruction results, and the F1 advantage of structured descriptions is observed across all three evaluated LLMs. These findings show that the appropriate representation depends on the target task and its required level of relational interpretation.
