Goal: Accurate vertebrae detection in 3D MR imaging is a fundamental preprocessing step for vertebra-level spine analysis, but remains challenging due to anisotropic resolution and inter-vertebral appearance similarity. Existing one-stage detectors lack explicit anatomical modeling, while multi-stage pipelines suffer from error propagation. Methods: We propose a Hybrid-Dimension Height-Aware Attention Network (HHANet), an end-to-end one-stage detector that couples anisotropic hybrid-dimension representation learning with anatomically guided height attention. A hybrid-dimension hierarchical extractor emphasizes within-slice representation while incorporating inter-slice context, and a height-attention detection head learns vertebra-specific height priors to enhance vertebral identification and localization. Results: Experiments on lumbar spine MR images demonstrate that HHANet outperforms state-of-the-art methods, achieving an Intersection over Union of 81.44% and an Identification Rate of 98.24%, with an inference time of 1.46×10-1 s per 320×160×16 MR volume Conclusions: HHANet enables reliable vertebra detection without multi-stage or sequential inference, providing an efficient foundation for downstream spine analysis.
Objective: Accurate characterization of arterial wall mechanics underlies vascular disease research and device design, yet ex vivo tensile testing relies on preconditioning protocols whose adequacy is often assumed rather than tested. This study fits a single-exponential length-decay model to cyclic stretch data from porcine carotid arteries to evaluate whether the commonly used 10-cycle standard is sufficient and how storage conditions influence arterial mechanical response. Results: Model-based analysis projects that arteries typically require approximately 30 loading cycles to reach a stable final length, exceeding the commonly used 10-cycle standard, with a mean fit quality of $R^{2} > 0.98$. Storage condition affected stabilization behavior for some samples. Conclusions: Widely used arterial preconditioning protocols may underestimate the number of cycles needed to reach a stable baseline, risking systematic overestimation of arterial stiffness in mechanical testing. Modeling cyclic stretch length decay offers a more robust, quantitatively validated framework for determining preconditioning requirements and improving consistency across studies.
Goal: Continuous health monitoring using wearables is increasingly important for the early detection of neurodegenerative disorders such as Alzheimer's disease. However, most EEG-based diagnostic systems rely on cloud processing and high-power hardware, making them unsuitable for long-term wearable deployment. Achieving continuous cognitive screening on resource-constrained sensing platforms requires energy-efficient signal processing, compact models, and minimal wireless communication. Methods: This paper presents an energy-efficient Edge-AI framework designed for microcontroller-class wearable devices. The system combines on-device preprocessing with a TinyML-optimized multi-scale Inception neural network to perform real-time inference directly on the sensing node. The pipeline incorporates an energy-aware Artifact Variance Gate to reject high-variance corrupted signal windows before inference while conserving computational energy. The variance gate serves as a hardware fail-fast mechanism that discards corrupted signal windows before inference. Next, we employ exponential moving standardization for adaptive signal normalization. Further, we introduce a variance-aware, learnable step-size quantization strategy that compresses the model to INT8 precision via quantization-aware training, enabling deployment within strict memory budgets. The model performs binary classification of EEG recordings from subjects with clinically diagnosed Alzheimer's disease (AD) and cognitively normal (CN) controls. Results: Evaluation on the OpenNeuro ds004504 dataset using Leave-One-Subject-Out (LOSO) Cross-Validation achieves 86.2% classification accuracy while reducing model memory requirements by $\sim 4\times$ compared with the original FP32 model. Conclusions: Finally, we developed a helmet prototype with a deployed quantized model embedded with edge hardware and evaluated it on an independent cohort of 20 subjects using a 19-channel wet-electrode.
This study evaluates whether isolating the extended cortico-striato-thalamo-cortical (CSTC) circuit improves Tourette Syndrome (TS) classification compared to whole-brain analysis using convolutional neural networks (CNNs). Analyzing 68 T1-weighted MRIs (34 TS, 34 controls), we compared FreeSurfer (DKT-Atlas) and SLANT segmentation methods. The pipeline included preprocessing, MONAI-based data augmentation, and classification via VGG16, VGG19, and ResNet50 with 5-fold cross-validation. Results show CSTC-based segmentation outperforms whole-brain models. VGG16 with FreeSurfer achieved the best performance (82.4% accuracy, 80.0% sensitivity, 83.0% specificity), surpassing deeper architectures that exhibited overfitting on the limited sample. SLANT segmentation followed closely with 80.9% accuracy. These findings confirm the CSTC circuit captures critical neuroanatomical signatures. By aligning with expert inter-rater reliability, this deep learning framework provides a proof-of-concept “biological second opinion” to support early TS diagnosis in pediatric neurology.
Low-cost, non-invasive screening tools for lumbar degenerative disease could complement imaging in resource-limited settings. This study presents a machine-learning framework to classify spine sounds recorded during lumbar motion and distinguish subjects with MRI-confirmed intervertebral disc and facet joint degeneration from asymptomatic controls. From each recording, we extract a 127-dimensional feature vector (MFCC statistics and temporal/spectral energy descriptors) and apply L1-norm-inducing kernel PCA (min and min-sign kernels) to obtain a compact, outlier-robust latent representation. Classification is performed using standard learners (SVM, Random Forest, Gradient Boosting, and KNN) under leave-one-patient-out (LOPO) cross-validation. The best configuration (min-kernel L1-kPCA with 20 components) achieves 93.3% accuracy with 93.0% sensitivity and 0.952 specificity (95% Wilson CI: 0.842–0.987) with an SVM classifier under rigorous LOPO cross-validation, outperforming the no-kPCA baseline (90.8% accuracy) and the neural network-based solution (93.3%). These results suggest that robust nonlinear feature compression can substantially improve spine sound classification and motivate external validation toward point-of-care screening.
