Guanjun Wang

Background: Speech holds promise for depressive symptom screening, but practical applications require clearer guidance on effective segmentation, optimal speech tasks, minimum duration, and feature-model matching. Objective: To enable low-burden screening for depressive symptoms, this study aims to evaluate the effectiveness of short speech samples and systematically investigate how to identify the optimal practical combination of different speech tasks, time-window lengths, and the matching between features and models. Methods: Using the HMU-DDAC-24 dataset of 143 participants aged 18–22 years, we evaluated 10 segmentation lengths (0.1–15 s), 5 models (Random Forest, Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Transformer, Extended Long Short-Term Memory (xLSTM)), and 3 feature types (frame-level, statistical, fused). We also assessed 19 window lengths (1–35 s) across four speech tasks and their combinations. In addition, supplementary cross-dataset validation analysis was performed using the external Emotional Audio-Textual Depression (EATD) dataset. Results: Frame-level features performed best with sequence models and short windows (0.1–2 s; Transformer accuracy: 0.731 at 2 s). Statistical features paired best with tree models and short windows (Random Forest accuracy: 0.698 at 1 s). Fused features improved tree models (Random Forest accuracy: 0.703 at 4 s) but introduced noise into sequence models. Task 4 (positive recall) achieved the best overall performance (Area Under the Receiver Operating Characteristic Curve (AUC) = 0.860), outperforming all single and combined tasks, showing no multi-task synergy. A 1-s segment performed comparably to full-length speech (3.14% relative performance loss). Kullback-Leibler divergence analysis showed short segments closely approximated the full recording’s distribution (Task 4 KL = 9.56 at 1 s). Conclusions: When combined with appropriate speech tasks and feature-model matching, short speech samples can facilitate low-burden screening for depressive symptoms. Supplementary analysis on the EATD dataset supported these trends, though absolute performance depended on feature representation. We recommend positive recall with 2-s truncation, frame-level features, and Transformer for accuracy-oriented settings; and 1-s truncation, statistical features, and Random Forest for resource-limited settings. These findings may inform rapid, remote speech-based screening for depressive symptoms, especially in resource-limited primary-care settings.

Background: Early screening for depressive symptoms lacks objective and stable biomarkers, while traditional methods based on questionnaires and subjective assessments suffer from issues such as low efficiency. Although the use of artificial intelligence to analyze speech signals has become a growing trend, existing methods are limited by challenges such as small medical datasets, complex features, and long-term temporal dependencies, resulting in insufficient recognition accuracy and poor cross-dataset generalization performance. Methods: To address the above issues, this paper proposes MAX-Net, a multi-scale feature-enhanced long-term sequential autoencoder network based on an improved extended long short-term memory (xLSTM) for depressive symptom screening. MAX-Net combines the improved xLSTM with autoencoder feature enhancement and introduces a Multi-Scale Convolutional layer (MSC) module. To mitigate the issue of data sparsity, we propose an autoencoder-based feature space enhancement strategy. Furthermore, we investigate the impact of varying numbers of augmented samples on model performance. The model is trained and evaluated on the in-house dataset HMU-DDAC-24 and retrained on the public dataset MODMA using the same architecture and hyperparameters to assess its stability and cross-dataset applicability. Results: Experiments on HMU-DDAC-24 show that MAX-Net achieves an Area Under the ROC Curve (AUC) of 0.945 and a Mean Absolute Error (MAE) of 1.61 in the depression symptom classification task, significantly outperforming existing baseline models. It maintains competitive performance on the MODMA dataset (AUC = 0.804), suggesting architectural stability and consistent effectiveness across different data sources when retrained. Further experiments revealed that the optimal data augmentation strategy varies across different datasets: on HMU-DDAC-24, performance peaks when 15,000 samples are generated, whereas on MODMA, performance stabilizes when 18,000–22,000 samples are generated. Conclusion: MAX-Net provides an effective framework for speech-based early depression screening and demonstrates consistent performance across datasets under a unified configuration. These findings indicate that the proposed architecture is robust to variations in data distribution. Furthermore, the autoencoder-based augmentation strategy has proven beneficial in low-data-volume scenarios. Overall, this study supports the application potential of MAX-Net as an auxiliary tool for early depression screening.

In the restoration of mural images with rich colour information and complex texture structures, existing techniques typically extract the spatial-domain features in the Red–Green–Blue (RGB) colour space. However, the three RGB channels are physically decoupled without unified perceptual colour correlation constraints, which often leads to noticeable colour deviation in damaged regions with large colour variations. In addition, restoring both high-frequency texture details and low-frequency global structures in a mixed-frequency spatial domain can create conflicts between frequencies, making it difficult to generate realistic high-frequency details. To address these issues, we propose the laboratory frequency network (LABFNet), a restoration network guided by the laboratory (LAB) colour space and frequency-domain constraints. Our model has two key improvements: (1) it incorporates colour parameters from the LAB space to model colour loss in murals, and (2) it decomposes the image into low- and high-frequency components and enforces frequency consistency during restoration. In the Dunhuang 20–40% mask-ratio setting, the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) improved by 1.58% and 0.27%, respectively, while the Mean Absolute Error (MAE), Learned Perceptual Image Patch Similarity (LPIPS) and CIEDE2000 decreased by 5.87%, 6.5%, and 21.65%, respectively. Experimental results on benchmark datasets show that LABFNet reduces colour deviation and structural defects.

Sperm morphology detection technology has important research value in diagnosing male infertility. Traditional manual detection methods are time-consuming, labor-intensive, and highly subjective. The sperm morphology analysis method based on deep learning can objectively and efficiently complete the sperm detection task, and has the advantage of end-to-end learning. This study proposed a two-stage sperm morphology analysis method, which used the UNet + + network to segment individual sperm, and then used the Multi-Head Mobilevit Net model based on multi-output classification to achieve one-time detection of the head, midpiece, and principal piece of a single sperm, and integrated the Mixing Loss function into the classification model. Based on the data enhancement technology, the expanded 10,802 improved Papanicolaou stain sperm morphology classification dataset was tested. The results showed that the accuracy of the sperm head, midpiece, and principal piece reached 83.47%, 96.19%, and 94.99%, respectively. This study can quickly realize the automated detection of sperm morphology, providing strong technical support for the development of reproductive medicine.

Fiber Bragg Grating (FBG) sensors are widely used for temperature monitoring in various applications through the demodulation of resonant wavelengths. Real-time demodulation of local high temperature variations is crucial for structural health monitoring, aero-engines, and related fields. However, local high-temperature characteristics of FBGs, such as resonance peak splitting and bandwidth increase, pose challenges to precise wavelength demodulation. To address these issues, this paper proposes a novel wavelength demodulation framework based on neural networks, employing an encoder-decoder architecture to achieve low-latency and high-precision results. Specifically, the reflectance spectrum, serving as the original sensing information, is encoded into different forms and processed through two branches. One branch encodes it as a Markov Transition Field and uses a pre-trained ResNet to extract underlying physical information, while the other utilizes a 1D convolutional neural network to extract wavelength properties. Finally, attention modules merge representations from the two branches, which are then decoded to accomplish wavelength demodulation. The effectiveness and superiority of our framework are validated on a real temperature monitoring system we built, achieving a minimum wavelength demodulation error and dynamic interrogation range of ±0.07 pm and 75 nm. These results highlight the potential of the framework for near-field temperature monitoring in complex environments.

The prediction of optical properties using neural network is a pivotal area in the design of photonic crystal fiber (PCF) structure. A critical challenge in this domain is the creation of a comprehensive training dataset, which is often impractical due to the inherent imbalance in optical property data. This paper proposes a novel framework aimed at enhancing the prediction of PCF optical properties, addressing the issue of data imbalance at minimal cost. Specifically, our framework leverages a Wasserstein Generative Adversarial Network with gradient penalties to augment the dataset by generating effective optical property data with high diversity from noise, thereby preserving the integrity of the original data. Following this, a specialized neural network is utilized to integrate both the original and augmented datasets, facilitating the extraction of broader optical property patterns. The efficacy and superiority of our proposed framework have been demonstrated through extensive qualitative and quantitative analyses on a Porous Core Single Mode Fiber structure. Following the implementation of our augmentation strategy, the mean-square error (MSE) for the optical properties prediction model decreased from 3.702×10⁻⁴ to 1.564 × 10⁻⁴ during training. Additionally, an R² value of 98.65% was achieved during testing. Our proposed framework provides a potential opportunity to achieve cost-effective structure and optimization of optical fiber structures.

Conventional fiber Bragg grating (FBG) accelerometer demodulation often suffers from high environmental sensitivity, complexity, and cost. To address these issues, this paper presents two arrayed waveguide grating (AWG)-based dynamic spectral reconstruction models. First, a novel differential FBG accelerometer with an equal-intensity triangular cantilever beam and a symmetric string-like design enhances sensitivity and mitigates temperature interference. A power ratio-nonlinear logarithmic regression (PR-NLR) model derived from static calibration then enables spectral reconstruction and wavelength demodulation across varying accelerations, achieving 4–82pm accuracy, slightly below that of high-speed demodulators (HSD). Next, by integrating convolutional neural networks (CNN), bidirectional long short-term memory (BiLSTM), attention mechanisms, and a generative adversarial network (GAN) for data augmentation, prediction errors drop to 1–26pm at 20Hz and 0.2–21pm at 40Hz-a reduction of at least 60% over PR-NLR. This outcome underscores the advantages of deep learning (DL) in managing complex spectral shifts. By overcoming frequency-only demodulation limits, the proposed AWG approach achieves precise wavelength demodulation through dynamic spectral reconstruction. Its compact design, low manufacturing cost, and robust performance facilitate scalable deployment in structural health monitoring, particularly for bridge vibration analysis and industrial machinery diagnostics.

A cascaded ultraviolet (UV) adhesive-based fiber-optic Fabry-Perot sensor utilizing the Vernier effect is proposed and experimentally validated. The sensor consists of two cascaded Fabry-Perot cavities, formed by fusing a section of single-mode fiber (SMF) to the interior of a short capillary glass tube (HST). A polymer cavity is created inside the capillary glass tube by initially introducing UV adhesive, followed by the formation of a series of polymer air cavities by filling the end faces of the capillary glass tube with a UV adhesive film. The sensor, benefiting from the Vernier effect, achieves a sensitivity on the order of -2.53 nm/kPa, exhibiting minimal temperature crosstalk. The pressure sensor demonstrates high sensitivity, cost-effectiveness, low resolution and excellent repeatability, rendering it a promising candidate for applications in medical and biological sensing.