
Objective: Nailfold capillary morphology provides clinically essential information for the noninvasive assessment of autoimmune, metabolic, and cardiovascular disorders. However, accurate segmentation of abnormal capillaries is highly challenging for conventional convolutional networks due to their thin, tortuous, and intertwined structures, as well as weak contrast against complex backgrounds. To address these limitations, we aimed to develop a novel network architecture explicitly designed for robust abnormal nailfold capillary segmentation.Approach:We developed TDS-U 2 Net, a novel variant of the U 2 -Net architecture. The proposed framework integrates dynamic snake convolution (DSConv) to enhance sensitivity to slender and curved local geometries, and incorporates a Transformer module to strengthen long-range dependency modeling and multiscale contextual fusion. Furthermore, a topological continuity constraint loss (TCLoss) function is employed alongside cross-entropy loss to penalize structural discontinuities and reduce breaks in thin tubular predictions.Main results: Evaluated on a curated nailfold capillary dataset, TDS-U 2 Net outperformed existing methods, achieving an accuracy of 98.33%, a Dice coefficient of 88.56%, an Intersection over Union of 80.92%, an F1-score of 88.23%, and the lowest Hausdorff distance of 6.98 pixels. Comprehensive ablation experiments confirmed that the inclusion of DSConv, the Transformer module, and TCLoss each contributed complementary performance improvements.Significance: By significantly improving structural continuity and preserving fine morphological details, the TDS-U 2 Net framework successfully overcomes the limitations of conventional segmentation networks. This approach provides a highly robust basis for downstream quantitative nailfold capillaroscopy analyses, facilitating the reliable extraction of clinical biomarkers such as vessel length, width, and density.
Automated radiographic grading of knee osteoarthritis is summarized by accuracy or quadratic weighted kappa (QWK), which can hide failure at Kellgren–Lawrence (KL) Grade 1. We studied this early-grade collapse using ImageNet-pretrained EfficientNet-B4 models (380 pixels) under a fixed protocol on the official Knee Osteoarthritis Dataset with Severity split. File names follow a seven-digit-identifier-plus-side pattern consistent with participant identifiers; interpreted this way, the split is subject-disjoint (4130 identifiers, both knees in the same split). A five-seed 2 × 2 factorial design crossed weighted cross-entropy with mild class-balanced sampling; a five-seed paired comparison evaluated CORAL-style ordinal regularization with pairwise ranking. QWK, the primary endpoint, did not differ detectably between configurations (0.824–0.838). Plain cross-entropy recalled only 0.118 ± 0.068 of Grade-1 knees. Class weighting and sampling raised Grade-1 recall (main effects +0.091 and +0.083; Holm-adjusted p = 0.016 and 0.029), reducing Grade-1 specificity (−0.034 and −0.029). Threshold-free Grade-1 discrimination was unchanged (average precision 0.350–0.360; AUROC 0.755–0.763), and a validation-selected Grade-1 decision offset applied to plain cross-entropy achieved comparable precision–recall operating points. Ordinal regularization improved no endpoint over the imbalance-aware baseline (QWK difference −0.005, 95% CI −0.019 to 0.008); ordinal weight had no detectable effect across three seeds, while blending ordinal-head probabilities monotonically lowered Grade-1 recall. Validation-tuned class-bias correction improved validation accuracy in all six runs but reduced test accuracy in four, left it unchanged in one, and improved it in one (mean test change −0.006). Retraining after removing perceptual-hash near-duplicate candidates did not change performance. Within this benchmark, early-grade collapse is largely a decision-threshold phenomenon: imbalance handling shifts the Grade-0/Grade-1 operating point without improving Grade-1 separability. Grade-1 sensitivity and specificity should accompany aggregate agreement.
Spondylolisthesis and degenerative disc disease can, in some cases, be treated using lumbar fusion techniques; however, the effects of these procedures on the surgical site, adjacent vertebral bodies, and intervertebral discs are not fully understood. In this work, a three-dimensional finite element model of the L1-L5 lumbar spine was developed using Computed Tomography data from a healthy individual, to evaluate stress redistribution following L4–L5 fusion. Four models were evaluated: a healthy spine without instrumentation; a spine with pedicle screw fixation at L4-L5; a spine with pedicle screw fixation and interbody cages at L4-L5; and a spine with interbody fusion at L4-L5. The analysis considered three loading conditions: axial compression, flexion, and extension. Percentage changes in vertebrae cortical-bone and intervertebral lumbar disc stress were calculated relative to the healthy-spine value under the same loading condition. Average von Mises stresses in the L1-L3 cortical bone and the L1-L2 and L2-L3 intervertebral discs remained similar to the healthy spine. Fusion reduced stress in the L4 vertebra relative to the screw-only construct, whereas L5 cortical bone stress increased across all instrumented configurations. Screw only fixation produced the largest increases, while the fused screw and cage construct produced the smallest, demonstrating that construct design and fusion state influence load redistribution at the fused level. The superior adjacent L3–L4 disc experienced a consistent 34–42% increase in stress across all instrumented models, indicating elevated adjacent segment loading even when cage-based constructs reduced stresses at L5. Flexion generated the highest stresses, approximately 90% greater than extension and 35% greater than compression, highlighting the sensitivity of postoperative load sharing to loading direction. These findings show that L4–L5 fusion alters stress distribution within the fused segment and the superior adjacent level, providing comparative biomechanical insight into how construct design and loading conditions influence postoperative load transfer.
Efficient P300 detection in brain-computer interfaces (BCIs) remains a challenge due to low signal-to-noise ratio (SNR) and the high computational cost of existing deep learning models, which limits their deployment in real-time applications. To address these issues, we propose MSTA-CNN, an ultra-lightweight multi-scale temporal-spatial attention convolutional neural network with only 3,153 trainable parameters. By synergizing depthwise separable convolutions for hierarchical feature extraction with a dual-branch attention mechanism designed to dynamically re-weight informative channels and temporal segments MSTA-CNN achieves robust end-to-end learning from minimally preprocessed EEG data. Extensive evaluations on the BCI Competition II Dataset IIb and III Dataset II demonstrate state-of-the-art character recognition accuracy, reaching 98% at 15 sequences and an average of 97.5% across subjects. Notably, the model yields a practical information transfer rate (ITR) of 13.72 bits/min, significantly outperforming baseline methods in critical low-repetition scenarios while being 10-50× lighter. The learned representations align closely with P300 neurophysiology, as evidenced by spatial importance maps and t-SNE visualizations. MSTA-CNN establishes a new benchmark for efficient P300 spellers, paving the way for portable, high-speed BCI systems for individuals with severe motor disabilities.
Numerical computations of induced electric field strength E i inside the human body are prone to so-called “staircasing errors”, i.e., physically incorrect local enhancements of the computed E i ate such numerical artifacts, ICNIRP suggested in their 2010 guidelines to use the 99 th percentile of the distribution of the 2 x 2 x 2 mm 3 at the edges of the cubical cells. Similar artifacts can occur due to imperfections in the anatomical body models used. In order to elimin averaged induced electric field strength Ei,avg for comparison with the basic restrictions instead of the maximum computed E i,avg value. However, this suggestion was based on very early computational studies using anatomical body models with relatively low spatial resolution. Moreover, a 99 th percentile-based assessment of E i,avg appears generally questionable in case of highly localized exposure scenarios, where large fractions of the body remain almost unexposed, and therefore true local hot spots of Ei,avg that should be considered may be excluded from the analysis, simply because the exposed tissue fraction is less than 1 % of the total body volume. Although this problem has been repeatedly pointed out by several research groups in the past, the 99 th percentile-based assessment unfortunately spread out as a “general rule” into some guidelines for exposure assessment. Consequently, the compliance of many devices on the market with exposure limits can be questioned. In this paper we are quantifying the potential underestimation of exposure when using the 99 th percentile-based Ei,avg for several representative scenarios of a highly localized magnetic field exposure of the hand using a state-of-the-art anatomical body model. Besides comparing the 99 th percentile-based Ei,avg values with values obtained when using higher percentiles, we also compare with E i,avg values obtained by applying a statistics-based outlier elimination method proposed by Arduino et al. (2020). Our results clearly demonstrate the inappropriateness of the 99th percentile-based assessment, as it may lead to underestimations of the actual induced E i,avg by more than a factor of 100, depending on the exposure scenario. In contrast, the statistics-based artifact elimination strategy proves to be a reasonable and robust method for estimating the actual value Ei,avg for highly localized magnetic field exposure of the hand.