Optical imaging technology through dynamic and complex scattering media has found applications in various fields, ranging from scene imaging to cell imaging. However, affected by inhomogeneous scattering, the imaging process in complex media faces numerous challenges, which can induce nonlinear effects. Although some studies have attempted to address the nonlinear effects in partial dynamic and complex scattering scenarios, this often requires a high sampling rate, and an efficient solution for this problem is still needed.This study proposes a Multi-Scale Convolutional Attention U-Net (MCAGI) method, which integrates an Adaptive Moving Average (AMA) correction module and an Efficient Multi-Scale Convolutional Attention Decoder (EMCAD). Specifically: the AMA can eliminate the interference of scaling factors caused by scattering; the EMCAD can focus on key features at low sampling rates while reducing computational costs.The results of simulation experiments and dynamic scattering experiments confirm that the MCAGI+AMA scheme can still reconstruct clear images at a 32% sampling rate, and its Peak Signal-to-Noise Ratio (PSNR) is 2 dB higher than that of the AMA+GIDC scheme. This study will provide support for expanding the application of Ghost Imaging (GI) in complex scattering environments.

Accurate assessment of pathological complete response (pCR) after neoadjuvant therapy is clinically important in breast cancer, yet reliable preoperative evaluation remains challenging. Serum vibrational spectroscopy provides a label-free and minimally invasive approach for capturing biochemical variation associated with subsequent treatment response in biofluids. However, clinical application is often constrained by limited sample size, class imbalance, and incomplete acquisition of multimodal spectral data. To address these constraints, we developed an availability-aware multimodal spectroscopy and mask-aware learning framework for serum pCR prediction, combining 532 nm Raman, 785 nm surface-enhanced Raman spectroscopy (SERS), and Fourier transform infrared (FTIR) spectra with CGAN-based augmentation and adaptive multimodal fusion. A conditional generative adversarial network (CGAN) was used to generate synthetic spectra for training augmentation, while MRAM-Net adaptively fused available spectral representations without spectral imputation. Using repeated patient-level cross-validation, the framework achieved an AUC of 94.2% when all modalities were available and maintained robust performance when one modality was unavailable (AUC >88%). In the independent validation cohort, the locked model achieved an AUC of 95.8% and an accuracy of 92.3% using the development threshold. These results indicate that the availability-aware framework can maintain stable performance under clinical data constraints. The proposed framework provides a practical strategy for treatment response assessment in breast cancer and a methodological basis for analyzing incomplete multimodal vibrational spectroscopy datasets in translational applications.

At present, Raman spectroscopy combined with deep learning has been widely used in the field of disease screening. Transformer is an important architecture for deep learning and has excelled in several areas with technologies such as its self-attention mechanism. However, as an architecture originally designed for the field of natural language processing, Transformer has disadvantages such as high computational complexity and easy overfitting in small data sets when processing spectral data. In this study, we propose a spectral classification model called Categorical Embedding Transformer (CET) and apply it to the screening of breast cancer and ductal carcinoma in situ combined with Raman spectroscopy. The core principle of CET model is to embed class labels to fixed dimensional vectors and update them as learnable parameters during training. The CET model also removes the positional encoding in transformer encoder and the initial linear layer used for dimensionality reduction or dimensionality enhancement, and retains the structure used for feature extraction and dimensionality reduction of spectral data. The ability of feature extraction and dimensionality reduction of spectral data is retained while the computational complexity is reduced. Finally, the dot product is used to calculate the similarity between the class vector and the spectrum after dimensionality reduction, and the cross entropy loss function is used to maximize the dot product similarity of the real class during training. The model we built achieved 100 % accuracy on the validation set and 98.2 % accuracy on the unknown test set, which is better than other compared models.

Cancer poses a serious threat to global public health, and early and accurate diagnosis is crucial for improving patient prognosis and reducing mortality rates. In this study, we used serum Raman spectroscopy combined with PCA-CNN model to diagnose three types of liver diseases, including hepatitis B, liver cirrhosis, and Hepatocellular Carcinoma. Established a serum Raman spectroscopy database containing three types of diseases and healthy controls. Compared with traditional methods such as SVM and PCA-LDA, PCA-CNN has higher efficiency and accuracy. Based on the collected 327 samples, PCA was used to extract features with a 95 % contribution rate and the 1D-CNN model was used for four classification. After five fold cross validation, the overall weighted average accuracy was 83.79 %. When the precision was 0.85,specificity was 0.95, F1 was 0.84, and AUC was 0.94, it improved by at least 4 % compared to traditional classification methods such as SVM. It provides a new method for rapid and non-invasive identification of multiple selected liver diseases, and contributes to clinical diagnosis.

Hepatitis B (HB) and hepatitis B related liver cirrhosis (LC) severely threaten global public health. Early and accurate differentiation of HB, LC, and healthy controls (HE) is crucial for optimizing treatment and improving prognosis. This study demonstrated that serum Raman spectroscopy combined with support vector machine (SVM) enables rapid diagnosis of HB and HB-induced LC. We established a Raman spectral database of 78 HB cases, 73 LC cases, and 81 HE cases, adopting a nested fivefold cross-validation (fivefold CV) scheme: Hyperparameters were tuned via grid search (internal fivefold CV) on the training set, with final performance estimated on an independent external test set. The entire process was repeated 10 times with different random seeds, and results were macro-averaged. The three-group classification achieved an overall accuracy of 78% and a macro-average area under the curve (AUC) of 91%, outperforming random forest (RF), principal component analysis-linear discriminant analysis (PCA-LDA), and k-nearest neighbors (KNN). This study confirms the method's great potential for noninvasive screening of HB and LC.

Based on the characteristics of spectral data, Nabil Ibtehaz et al. (2023) proposed a generalized neural network architecture for Raman spectroscopy analysis, called RamanNet. This paper applies it to breast cancer screening and proposes an modified RamanNet method to optimize the classification performance of breast cancer and healthy individuals. The modified model accelerates convergence and reduces overfitting by incorporating L2 regularization, removing TripletLoss, and adjusting the learning rate. Results demonstrate that the modified RamanNet achieved a higher accuracy (96.0 f 1.7%) and sensitivity (96.8 f 3.0%) in distinguishing between breast cancer patients and healthy controls, outperforming both the 1D-CNN (accuracy: 91.8 f 2.9%; sensitivity: 89.3 f 5.1 %) and the original RamanNet (accuracy: 92.5 f 3.2%; sensitivity: 94.6 f 5.6 %). Furthermore, the model demonstrated enhancements in training time, convergence speed and stability, which provides a new technological approach for non-invasive and rapid breast cancer screening with great potential for clinical application.

We have proposed a direct computational imaging method via speckle patterns based on a multi society genetic algorithm. Decomposing the reconstruction problem of computational ghost imaging into two objectives, we introduce the multi-society genetic algorithm, enabling the reconstructed object image to be updated in the form of speckle patterns. The results demonstrate that this method can achieve a high quality reconstructed image. Compared with existing methods, our approach can improve the Peak Signal-to-Noise Ratio (PSNR) by up to 4.6 dB in experimental settings. This is beneficial for promoting biomedical imaging and developing integrated single pixel cameras.

Ghost imaging (GI) is capable of reconstructing images under low-light conditions by single-pixel measurements. However, improving image resolution often requires extensive single-pixel sampling, limiting practical applications. Here we propose a super-resolution algorithm of GI using Convolutional neural network with Grouped orthonormalization algorithm Constraint (GICGC), which aims to reconstruct images at super-resolution with strong local regularities and self-similarity. The proposed algorithm, a versatile approach, has been demonstrated to outperform several other widely used GI algorithms in terms of spatial resolution and sampling rate. Our findings are supported by rigorous benchmark tests and experimental validations in challenging environments, including multimode fibers. The imaging experimental results demonstrate that GICGC achieves superior performance in reconstructing image linewidth and contrast, effectively doubling resolution capability by approximately 2.1 times, indicating significant potential in biomedical imaging and other fields. We believe this study represents a novel breakthrough in universal super-resolution ghost imaging and paves the way for its practical applications.

The orbital angular momentum (OAM) of beams has important applications in various fields such as large-capacity data transmission, imaging, and quantum information. So the precise measurement of the OAM spectrum is critical. Here, we present a direct and efficient method for measuring the OAM spectrum in a scattering medium using a Faraday atomic filter. Experimental results show that the Faraday atomic filter offers significant advantages in accurately identifying OAM under scattering media when compared to the results in the filterless device and the polarizer. Our work presents a novel, to the best of our knowledge, approach to precisely measure the OAM spectrum of vortex beams in complex scattering environments and will contribute to the advancement of various applications, including optical communications, lidar, and others.

In moving target ghost imaging, one of the main challenges is the issue of sampling rate. To address this, this paper proposes an end-to-end method for reconstructing moving targets under low sampling conditions and has conducted relevant simulations. Based on the simulations of this method, the paper combines it with the Faraday Anomalous Dispersion Optical Filter (FADOF) to achieve moving target ghost imaging experiments under daylight noise conditions. Under conditions where the moving speed reaches 15 mm/s (with a distance of 60 mm between the object and the light source) and the number of sampling frames is 320, our simulation results still achieve a signal-to-noise ratio (SNR) of 7 dB, while the experimental results exceed 4 dB, which is more than 2 dB higher than the results of MTGI(Moving Target Ghost Imaging). This provides a reference for practical applications of ghost imaging for moving targets under sunlight.

In our research, we conducted a detailed analysis of the impact and causes of argon (Ar) as a buffer gas on the transmission characteristics of K-FADOF at the wavelenght of 767 nm and 770 nm. The experimental findings indicate that the incorporation of Ar has a profound effect on reshaping the spectral transmission of both the D1 and D2 lines. Notably, as the argon pressure rises, the sidebands of the D2 line are significantly subdued, while intriguingly, its central peak retains a transmission level of 40 %. Simultaneously, there is a general degradation in the transmission efficiency of the D1 line, which can be attributed to its formation mechanism mirroring that of the D2 line's sidebands, making it susceptible to suppression by argon. The strategy of introducing argon to modulate the transmission spectrum not only provides valuable insights for the design of high-performance 770 nm K-FADOF but is also applicable to Na-FADOF, aiding in the design of high-performance filters that can effectively suppress mutual interference caused by the mere 0.6 nm wavelength difference between their yellow doublets.

This research endeavors to boost the image quality of underwater computational ghost imaging systems amidst the challenges posed by background illumination and turbulent disturbances. Utilizing FADOF, we have empirically validated their advantage in mitigating interference and bolstering the robustness of imaging. comparison to traditional filters, FADOF has proven to be remarkably resilient and stable in conditions rife with interference. Additionally, the strategic infusion of argon gas into the FADOF as a buffer has refined its transmission capabilities, adeptly filtering out extraneous frequency signals. Quantitative evaluations indicate that our enhanced FADOF markedly improves the quality of reconstructed images, achieving a commendable equilibrium between imaging proficiency and resource expenditure, especially at a 5 torr fill level. The findings this study chart new courses for the evolution of underwater ghost imaging techniques, with far-reaching implications for augmenting our capacity for marine observation and fostering the sustainable exploitation marine resources.

LiDAR detection is very sensitive to the surface roughness of objects, which will affect the detection and identification results of the objects. The orbital angular momentum (OAM) spectrum can also be used with LiDAR to detect objects, but there are currently no research papers analyzing the impact of roughness in this scenario. In this study, a roughness model and a projection measurement method were used for numerical simulation. The analysis showed that when the wavelength is 0.5 um <= lambda <= 1.5 um, and the correlation length is 1 mm, roughness does not affect the rotational symmetry of objects identified the OAM spectrum. When the correlation length is 5 mm and 10 mm, and the root mean square of the surface roughness is alpha < 0.7 lambda, the roughness does not affect the determination of the rotational symmetry of the target. Experimentally, this study built an optical architecture for detecting the symmetry of objects using OAM spectra. Using a 532 nm laser, the study measured five-leaf clover objects with four different roughness levels (320 mesh, 600 mesh, 2000 mesh, and 10,000 mesh). The experimental results show that using OAM spectra to detect the rotational symmetry of objects is insensitive to surface roughness, which is consistent with the simulation results. This study can promote the further application of orbital angular momentum in LiDAR target recognition.

Lupus nephritis (LN) is one of the most common and serious organ manifestations of systemic lupus erythematosus (SLE), with a poor long-term prognosis and a complex diagnostic process, therefore it is important to find a simple, rapid and non-invasive method for the diagnosis of LN. This study investigated the feasibility of using surface-enhanced Raman spectroscopy (SERS) and Fourier Transform Infrared (FT-IR) spectroscopy of urine samples to classify healthy volunteers and LN patients. SERS and FT-IR data of urine samples were obtained from 100 LN patients and 100 healthy volunteers. To verify the stability of the classification algorithm, 50 independent experiments were conducted. In each experiment, the dataset was randomly divided and a classification model was established using the support vector machine (SVM) algorithm (linear kernel function). Meanwhile, it was compared with four other common classification algorithms and the results showed that SVM model had the best effect. The average classification accuracy of SERS and FT-IR spectra combined with SVM model for 50 independent experiments reached 96.97 % and 92.77 %, respectively. In addition, the features of SERS and FT-IR were spliced and then combined with SVM model for classification, corresponding to an average classification accuracy of 97.63 %. Subsequently, genetic algorithm was used to perform feature selection on the spliced features, and the 16 selected features were also input into SVM model, with an average classification accuracy of 99.47 % over 50 independent experiments. Therefore, urine vibrational spectroscopy combined with SVM model has great potential in the diagnosis of lupus nephritis.

Breast cancer is one of the most common tumors in women, and early screening can significantly reduce mortality rates. Meanwhile, accurately identifying HER2-positive and HER2-negative subtypes of breast cancer is critical for helping doctors determine treatment options and prognosis strategies for patients. The goal of this study was to develop a computationally efficient, end-to-end model capable of both breast cancer detection and molecular typing prediction without the need for complex feature engineering. This study collected serum samples from 541 volunteers, including HER2-positive, HER2-negative, ductal carcinoma in situ (DCIS) patients, and healthy individuals. After sample collection, Raman spectra were obtained using a Raman spectrometer with a 532 nm excitation wavelength. Based on an efficient channel attention mechanism and convolutional neural networks, a classification model was developed to facilitate breast cancer detection and molecular subtyping. The proposed model significantly reduced the number of parameters and increased training speed. On an unknown test set, the model achieved an accuracy of 94.5 % and an AUC of 0.952, outperforming traditional models and algorithms. Specifically, the model achieved an accuracy of 98.1 % for BC, DCIS, and healthy volunteers, and 89.5 % for HER2-positive and HER2-negative cases. Additionally, the model's effectiveness was validated using different datasets, yielding satisfactory results. Raman spectroscopy, combined with our proposed attention mechanism-based convolutional neural network, effectively enables early screening and molecular subtype prediction of breast cancer. These findings offer new possibilities for rapid, non-invasive, and low-cost early screening and molecular subtype prediction of breast cancer.

Computational ghost imaging (CGI) has emerged as a promising technique for diverse imaging applications, particularly in challenging environments. However, achieving high-quality image reconstruction under low sampling rates and noisy conditions remains a significant challenge hindering practical deployment. To overcome these limitations and achieve superior reconstruction quality, we present a novel CGI method based on the Alternating direction method of multipliers (ADMM) fused with Low-rank regularization (GIAL). We also develop a fiber-based ghost imaging setup for experimental validation. Numerical simulations and experimental results validate the exceptional and general reconstruction performance of the proposed GIAL algorithm. Our findings demonstrate the algorithm's remarkable capacity to reconstruct high-quality images at extremely low sampling rates (e.g., 1.56%) and highlight its inherent robustness to noise. These superior characteristics underscore the significant potential of the GIAL method for widespread applications in biomedical imaging and remote sensing scenarios. (To foster transparency and reproducibility, the complete implementation of GIAL is available at https://gitee.com/dlammm2066/GIAL, subject to journal policy).

Ghost imaging facilitates image reconstruction under low-light conditions, but traditionally requires a large number of samplings for high-quality reconstruction, which limits its practical applications. To address these challenges, we propose an unsupervised deep learning neural network for ghost imaging (UDNGI) that combines a physical forward model with an optimized U2-Net architecture, eliminating the need for pre-training, and is enhanced by a convolutional block attention module (CBAM) attention mechanism. We validated the proposed UDNGI in a ghost imaging system with multimode fiber. Experimental results show that UDNGI achieves an SSIM of approximately 0.43 at an ultra-low sampling rate of 0.008 and improves the resolution by 1.53 times at a sampling rate of 0.064. This advancement offers a promising solution for high-performance ghost imaging.

This study investigates the effects of underwater turbulence on computational ghost imaging (CGI) technology, combining theoretical analysis with experimental data. We simulated vehicle-induced turbulence and utilized compressed sensing for image reconstruction to analyze its impact on image quality. Our findings reveal that turbulence, especially flow-aligned with the light beam, significantly degrades image quality, with the most pronounced effects at the signal reception end. As turbulence increases, image quality declines, and turbulence directionality substantially affects image quality, with parallel turbulence causing more wavefront aberrations. CGI shows tolerance to light-perpendicular turbulence, maintaining clarity despite resolution, contrast, and brightness reductions during intense turbulence. We introduced PSNR and SSIM to quantify the impact of turbulence on image quality, confirming CGI's resilience to turbulence interference. Our results imply that optimizing the vehicle platform's propulsion system to minimize turbulence in the signal detection direction can preserve ghost imaging performance. This research lays a scientific groundwork for improving CGI's turbulence resistance in underwater environments, aiding its practical use in complex underwater scenarios.

Computational ghost imaging faces significant performance degradation in dynamic oceanic turbulence. This study proposes a 'orthogonal basis + controllable randomization' principle to design robust speckle patterns. We establish a multi-level evaluation system to analyze six types of speckles, revealing a dual mechanism: orthogonality suppresses mode degradation while randomization enhances turbulence adaptability. Results demonstrate that the synergistic design (e.g. orthogonal random speckle) significantly outperforms classical orthogonal speckles in maintaining structural integrity and suppressing crosstalk under strong turbulence. Furthermore, different randomization strategies (global vs block-wise) offer adaptable performance trade-offs between stability and fidelity. This work addresses the lack of dynamic optimization criteria in speckle design, providing a potential solution for high-resolution underwater imaging.

The quality and safety of pharmaceuticals are critical issues that directly impact human health. Developing rapid and cost-effective drug detection technologies can effectively ensure pharmaceutical quality and safety. Infrared spectroscopy, with its advantages of non-destructive testing, rapid response, and low cost, has gradually gained attention in the field of pharmaceutical analysis. This study proposes a dual-branch convolutional neural network integrated with an efficient channel attention mechanism (ECA) to classify the infrared spectra of escitalopram tablets at different dosages. The constructed model utilizes two convolutional branches with different parameters to comprehensively extract feature information from the input data, while the ECA enhances the model's ability to capture key information. On an unseen test set, the model achieved accuracy, precision, recall, and F1 score values of 98.81 %, 98.53 %, 98.44 %, and 98.44 %, respectively, outperforming traditional machine learning algorithms and one-dimensional CNNs. Experimental results demonstrate that the combination of deep learning and infrared spectroscopy holds great potential for rapid and cost-effective pharmaceutical detection.