Deep learning

Explainable Machine Learning Models for Prediction of Surgical Site Infection After Posterior Lumbar Fusion Surgery Based on Shapley Additive Explanations

OBJECTIVE

This study aims to develop machine learning (ML) models combined with an explainable method for the prediction of surgical site infection (SSI) after posterior lumbar fusion surgery.

METHODS

In this retrospective, single-center study, a total of 1016 consecutive patients who underwent posterior lumbar fusion surgery were included. A comprehensive dataset was established, encompassing demographic variables, comorbidities, preoperative evaluation, details related to diagnosed lumbar disease, preoperative laboratory tests, surgical specifics, and postoperative factors. Utilizing this dataset, 6nullML models were developed to predict the occurrence of SSI. Performance evaluation of the models on the testing set involved several metrics, including the receiver operating characteristic curve, the area under the receiver operating characteristic curve, accuracy, recall, F1 score, and precision. The Shapley Additive Explanations (SHAP) method was employed to generate interpretable predictions, enabling a comprehensive assessment of SSI risk and providing individualized interpretations of the model results.

RESULTS

Among the 1016 retrospective cases included in the study, 36 (3.54%) experienced SSI. Out of the six models examined, the Extreme Gradient Boost model demonstrated the highest discriminatory performance on the testing set, achieving the following metrics: precision (0.9000), recall (0.8182), accuracy (0.9902), F1 score (0.8571), and area under the receiver operating characteristic curve (0.9447). By utilizing the SHAP method, several important predictors of SSI were identified, including the duration of indwelling jugular vein catheter, blood urea nitrogen levels, total protein levels, sustained fever, creatinine levels, triglycerides levels, monocyte count, diabetes mellitus, drainage time, white blood cell count, cerebral infarction, estimated blood loss, prealbumin levels, Prognostic Nutritional Index, low back pain, posterior fusion score, and osteoporosis.

CONCLUSIONS

ML-based prediction tools can accurately assess the risk of SSI after posterior lumbar fusion surgery. Additionally, ML combined with SHAP could provide a clear interpretation of individualized risk prediction and give physicians an intuitive comprehension of the effects of the model's essential features.

Predicting the Degree of Fresh Tea Leaves Withering Using Image Classification Confidence

Rapid and non-destructive detection methods for the withering degree of fresh tea leaves are crucial for ensuring high-quality tea production. Therefore, this study proposes a fresh tea withering degree detection model based on image classification confidence. The moisture percentage of fresh tea leaves is calculated by developing a weighted method that combines confidence levels and moisture labels, and the degree of withering is ultimately determined by incorporating the standard for wilted moisture content. To enhance the feature extraction ability and classification accuracy of the model, we introduce the Receptive-Field Attention Convolution (RFAConv) and Cross-Stage Feature Fusion Coordinate Attention (C2f_CA) modules. The experimental results demonstrate that the proposed model achieves a classification accuracy of 92.7%. Compared with the initial model, the detection accuracy was improved by 0.156. In evaluating the predictive performance of the model for moisture content, the correlation coefficients (Rp), root mean square error (RMSEP), and relative standard deviation (RPD) of category 1 in the test set were 0.9983, 0.006278, and 39.2513, respectively, and all performance were significantly better than PLS and CNN methods. This method enables accurate and rapid detection of tea leaf withering, providing crucial technical support for online determination during processing.

Interfacing with the Brain: How Nanotechnology Can Contribute

Interfacing artificial devices with the human brain is the central goal of neurotechnology. Yet, our imaginations are often limited by currently available paradigms and technologies. Suggestions for brain-machine interfaces have changed over time, along with the available technology. Mechanical levers and cable winches were used to move parts of the brain during the mechanical age. Sophisticated electronic wiring and remote control have arisen during the electronic age, ultimately leading to plug-and-play computer interfaces. Nonetheless, our brains are so complex that these visions, until recently, largely remained unreachable dreams. The general problem, thus far, is that most of our technology is mechanically and/or electrically engineered, whereas the brain is a living, dynamic entity. As a result, these worlds are difficult to interface with one another. Nanotechnology, which encompasses engineered solid-state objects and integrated circuits, excels at small length scales of single to a few hundred nanometers and, thus, matches the sizes of biomolecules, biomolecular assemblies, and parts of cells. Consequently, we envision nanomaterials and nanotools as opportunities to interface with the brain in alternative ways. Here, we review the existing literature on the use of nanotechnology in brain-machine interfaces and look forward in discussing perspectives and limitations based on the authors' expertise across a range of complementary disciplines─from neuroscience, engineering, physics, and chemistry to biology and medicine, computer science and mathematics, and social science and jurisprudence. We focus on nanotechnology but also include information from related fields when useful and complementary.

Dynamic, Interpretable, Machine Learning-Based Outcome Prediction as a New Emerging Opportunity in Acute Ischemic Stroke Patient Care: A Proof-of-Concept Study

Introduction: While the machine learning (ML) model's black-box nature presents a significant barrier to effective clinical application, the dynamic nature of stroke patients' recovery further undermines the reliability of established predictive scores and models, making them less suitable for accurate prediction and appropriate patient care. This research is aimed at building and evaluating an interpretable ML-based model, which would perform outcome prediction at different time points of patients' recovery, giving more secure and understandable output through interpretable packages. Materials and Methods: A retrospective analysis was conducted on acute ischemic stroke (AIS) patients treated with alteplase at the Neurology Clinic of the University Clinical Center of Vojvodina (Novi Sad, Serbia), for 14 years. Clinical data were grouped into four categories based on collection time-baseline, 2-h, 24-h, and discharge features-serving as inputs for three different classifiers-support vector machine (SVM), logistic regression (LR), and random forest (RF). The 90-day modified Rankin scale (mRS) was used as the outcome measure, distinguishing between favorable (mRS ≤ 2) and unfavorable outcomes (mRS ≥ 3). Results: The sample was described with 49 features and included 355 patients, with a median age of 67 years (interquartile range (IQR) 60-74 years), 66% being male. The models achieved strong discrimination in the testing set, with area under the curve (AUC) values ranging from 0.80 to 0.96. Additionally, they were compared with a model based on the DRAGON score, which showed an AUC of 0.760 (95% confidence interval (CI), 0.640-0.862). The decision-making process was more thoroughly understood using interpretable packages: Shapley additive explanation (SHAP) and local interpretable model-agnostic explanation (LIME). They revealed the most significant features at both the group and individual patient levels. Conclusions and Clinical Implications: This study demonstrated the moderate to strong efficacy of interpretable ML-based models in predicting the functional outcomes of alteplase-treated AIS patients. In all constructed models, age, onset-to-treatment time, and platelet count were recognized as the important predictors, followed by clinical parameters measured at different time points, such as the National Institutes of Health Stroke Scale (NIHSS) and systolic and diastolic blood pressure values. The dynamic approach, coupled with interpretable models, can aid in providing insights into the potential factors that could be modified and thus contribute to a better outcome.

Transforming Agricultural Waste from Mediterranean Fruits into Renewable Materials and Products with a Circular and Digital Approach

The Mediterranean area is one of the major global producers of agricultural food. However, along the entire supply chain-from farming to food distribution and consumption-food waste represents a significant fraction. Additionally, plant waste residues generated during the cultivation of specific fruits and vegetables must also be considered. This heterogeneous biomass is a valuable source of bioactive compounds and materials that can be transformed into high-performance functional products. By analyzing technical and scientific literature, this review identifies extraction, composite production, and bioconversion as the main strategies for valorizing agricultural by-products and waste. The advantages of these approaches as well as efficiency gains through digitalization are discussed, along with their potential applications in the Mediterranean region to support new research activities and bioeconomic initiatives. Moreover, the review highlights the challenges and disadvantages associated with waste valorization, providing a critical comparison of different studies to offer a comprehensive perspective on the topic. The objective of this review is to evaluate the potential of agricultural waste valorization, identifying effective strategies while also considering their limitations, to contribute to the development of sustainable and innovative solutions in Mediterranean bioeconomy.

Deep learning image analysis for continuous single-cell imaging of dynamic processes in Plasmodium falciparum-infected erythrocytes

Continuous high-resolution imaging of the disease-mediating blood stages of the human malaria parasite Plasmodium falciparum faces challenges due to photosensitivity, small parasite size, and the anisotropy and large refractive index of host erythrocytes. Previous studies often relied on snapshot galleries from multiple cells, limiting the investigation of dynamic cellular processes. We present a workflow enabling continuous, single-cell monitoring of live parasites throughout the 48-hour intraerythrocytic life cycle with high spatial and temporal resolution. This approach integrates label-free, three-dimensional differential interference contrast and fluorescence imaging using an Airyscan microscope, automated cell segmentation through pre-trained deep-learning algorithms, and 3D rendering for visualization and time-resolved analyses. As a proof of concept, we applied this workflow to study knob-associated histidine-rich protein (KAHRP) export into the erythrocyte compartment and its clustering beneath the plasma membrane. Our methodology opens avenues for in-depth exploration of dynamic cellular processes in malaria parasites, providing a valuable tool for further investigations.

Deep-Learning-Assisted Understanding of the Self-Assembly of Miktoarm Star Block Copolymers

The self-assemblies of topological complex block copolymers, especially the ABn type miktoarm star ones, are fascinating topics in the soft matter field, which represent typical self-assembly behaviors analogous to those of biological membranes. However, their diverse topological asymmetries and versatile spontaneous curvatures result in rather complex phase separations that deviate significantly from the common mechanisms. Thus, numerous trial-and-error experiments with tremendous parameter space and intricate relationships are needed to study their assemblies. Herein, we applied deep learning technology to decipher the phase behaviors of the miktoarm star block copolymer PEO-s-PS2 in an evaporation-induced self-assembly system. A neural network model was trained from practical experimental data encompassing two polymer properties and three synthesis condition parameters as input variables, which successfully predicted a three-dimensional (3D) synthesis-field diagram and mined the relationship between input parameters and obtained structures. This model demonstrated the highly flexible structure modulation directions of the miktoarm star block copolymer, revealing the correlation between the polymer parameters, synthesis conditions, and the output structures due to the significant influence of the variables on spontaneous curvatures. This work demonstrated the efficiency of a deep learning technique in uncovering the underlying rules of complex self-assembly systems, providing valuable insights into the exploration of soft matter science.

IBPL: Information Bottleneck-based Prompt Learning for graph out-of-distribution detection

When training and test graph samples follow different data distributions, graph out-of-distribution (OOD) detection becomes an indispensable component of constructing the reliable and safe graph learning systems. Motivated by the significant progress on prompt learning, graph prompt-based methods, which enable a well-trained graph neural network to detect OOD graphs without modifying any model parameters, have been a standard benchmark with promising computational efficiency and model effectiveness. However, these methods ignore the influence of overlapping features existed in both in-distribution (ID) and OOD graphs, which weakens the difference between them and leads to sub-optimal detection results. In this paper, we present the Information Bottleneck-based Prompt Learning (IBPL) to overcome this challenging problem. Specifically, IBPL includes a new graph prompt that jointly performs the mask operation on node features and the graph structure. Building upon this, we develop an information bottleneck (IB)-based objective to optimize the proposed graph prompt. Since the overlapping features are inaccessible, IBPL introduces the noise data augmentation which generates a series of perturbed graphs to fully covering the overlapping features. Through minimizing the mutual information between the prompt graph and the perturbed graphs, our objective can eliminate the overlapping features effectively. In order to avoid the negative impact of perturbed graphs, IBPL simultaneously maximizes the mutual information between the prompt graph and the category label for better extracting the ID features. We conduct experiments on multiple real-world datasets in both supervised and unsupervised scenarios. The empirical results and extensive model analyses demonstrate the superior performance of IBPL over several competitive baselines.

Highly Accurate and Robust Early Stage Detection of Cholangiocarcinoma Using Near-Lossless SERS Signal Processing with Machine Learning and 2D CNN for Point-of-care Mobile Application

INTRODUCTION

Cholangiocarcinoma (CCA), a malignancy of the bile ducts, presents a significant health burden with a notably high prevalence in Northeast Thailand, where its incidence ratio is 85 per 100,000 population per year. The prognosis for CCA patients remains poor, particularly for proximal tumors, with a dismal 5-year survival rate of just 10%. The challenge in managing CCA is exacerbated by its typically late detection, contributing to a high mortality rate. Current screening methods, such as ultrasound, are insufficient, as many CCA patients do not exhibit prior symptoms or detectable liver fluke (Opisthorchis viverrini : OV) infections, underscoring the urgent need for alternative early detection methods.

METHODS

In this study, we introduce a novel approach utilizing surface-enhanced Raman spectroscopy (SERS) combined with near-lossless signal compression via discrete wavelet transform (DWT) together with 2D CNN for the first time. Hamster serums of different stages were collected as the data set. DWT was employed for feature extraction, enabling the capture of the entire SERS spectrum, unlike traditional methods like PCA and LDA, which focus only on specific peaks. These features were used to train a 2D convolutional neural network (2D CNN), which is particularly robust against translation, rotation, and scaling, thus effectively addressing the SERS peak shifting issues. We validated our approach using gold-standard histology, and notably, our method could detect CCA at an early stage. The ability to identify CCA at the early stage significantly improves the chances of successful intervention and patient outcomes.

RESULTS AND CONCLUSION

Our results demonstrate that our method, combining SERS with extremely compact wavelet feature extraction and 2D CNN, outperformed other approaches (PCA + SVM, PCA + 1D CNN, PCA + 2D CNN, LDA + SVM, and DWT + 1D CNN), achieving performance of 95.1% accuracy, 95.08% sensitivity, 98.4% specificity, and an area under the curve (AUC) of 95%. The trained model was further deployed on a server and mobile application interface, paving the way for future field experiments in rural areas and home-use potential point-of-care services.

Ultra-compact multi-task processor based on in-memory optical computing

To enhance the computational density and energy efficiency of on-chip neuromorphic hardware, this study introduces a novel network architecture for multi-task processing with in-memory optical computing. On-chip optical neural networks are celebrated for their capability to transduce a substantial volume of parameters into optical form while conducting passive computing, yet they encounter challenges in scalability and multitasking. Leveraging the principles of transfer learning, this approach involves embedding the majority of parameters into fixed optical components and a minority into adjustable electrical components. Furthermore, with deep regression algorithm in modeling physical propagation process, a compact optical neural network achieve to handle diverse tasks. In this work, two ultra-compact in-memory diffraction-based chips with integration of more than 60,000 parameters/mm2 were fabricated, employing deep neural network model and the hard parameter sharing algorithm, to perform multifaceted classification and regression tasks, respectively. The experimental results demonstrate that these chips achieve accuracies comparable to those of electrical networks while significantly reducing the power-intensive digital computation by 90%. Our work heralds strong potential for advancing in-memory optical computing frameworks and next generation of artificial intelligence platforms.

How AI-Driven Root and Bone Predictions Can Assist Clear Aligner Treatment Planning

Integrating artificial intelligence (AI) and advanced three-dimensional (3D) imaging has revolutionised dentistry by enhancing diagnostics and treatment planning. Advanced algorithms and machine-learning techniques may enable orthodontists to analyse complex cases and predict treatment outcomes accurately. This technology facilitates the creation of customised treatment plans that consider individual tooth morphology and periodontal health, optimising force application and minimising treatment time. Since their introduction, clear aligners have gained popularity, with over 17 million people treated by 2023. Compared with fixed appliances, clear aligners offer advantages, such as better aesthetics, comfort and oral hygiene. Treating patients with a compromised periodontium requires accurate diagnosis and treatment planning. This paper reviews how AI-driven treatment planning software predicting root movement and visualising bone structures may impact treatment decisions and, ultimately, treatment outcomes. The technology behind machine learning and AI in designing clear aligners is discussed. Research shows that when viewing the cases in 3D, clinicians are more comfortable when treating crowding cases with a non-extraction approach using interproximal reduction (IPR) only. However, it was interesting to note that clinicians with extensive experience treating clear aligner patients were more comfortable using IPR to address severe crowding cases when viewed in 2D, compared with those less experienced with clear aligners. However, when the cases were visualised in 3D, both groups showed equal comfort in using IPR, as the roots were within the bone. AI-driven treatment planning software, utilising machine learning in conjunction with 3D modelling, may enhance the predictability of orthodontic movements while reducing treatment time and increasing patient satisfaction.

A New Perspective on Precision Medicine: The Power of Digital Organoids

Precision medicine is a personalized medical model based on the individual's genome, phenotype, and lifestyle that provides tailored treatment plans for patients. In this context, tumor organoids, a 3-dimensional preclinical model based on patient-derived tumor cell self-organization, combined with digital analysis methods, such as high-throughput sequencing and image processing technology, can be used to analyze the genome, transcriptome, and cellular heterogeneity of tumors, so as to accurately track and assess the growth process, genetic characteristics, and drug responsiveness of tumor organoids, thereby facilitating the implementation of precision medicine. This interdisciplinary approach is expected to promote the innovation of cancer diagnosis and enhance personalized treatment. In this review, the characteristics and culture methods of tumor organoids are summarized, and the application of multi-omics, such as bioinformatics and artificial intelligence, and the digital methods of organoids in precision medicine research are discussed. Finally, this review explores the main causes and potential solutions for the bottleneck in the clinical translation of digital tumor organoids, proposes the prospects of multidisciplinary cooperation and clinical transformation to narrow the gap between laboratory and clinical settings, and provides references for research and development in this field.

Development of a deep learning-based model to evaluate changes during radiotherapy using cervical cancer digital pathology

This study aims to create a deep learning-based classification model for cervical cancer biopsy before and during radiotherapy, visualize the results on whole slide images (WSIs), and explore the clinical significance of obtained features. This study included 95 patients with cervical cancer who received radiotherapy between April 2013 and December 2020. Hematoxylin-eosin stained biopsies were digitized to WSIs and divided into small tiles. Our model adopted the feature extractor of DenseNet121 and the classifier of the support vector machine. About 12 400 tiles were used for training the model and 6000 tiles for testing. The model performance was assessed on a per-tile and per-WSI basis. The resultant probability was defined as radiotherapy status probability (RSP) and its color map was visualized on WSIs. Survival analysis was performed to examine the clinical significance of the RSP. In the test set, the trained model had an area under the receiver operating characteristic curve of 0.76 per-tile and 0.95 per-WSI. In visualization, the model focused on viable tumor components and stroma in tumor biopsies. While survival analysis failed to show the prognostic impact of RSP during treatment, cases with low RSP at diagnosis had prolonged overall survival compared to those with high RSP (P = 0.045). In conclusion, we successfully developed a model to classify biopsies before and during radiotherapy and visualized the result on slide images. Low RSP cases before treatment had a better prognosis, suggesting that tumor morphologic features obtained using the model may be useful for predicting prognosis.

iScore: A ML-Based Scoring Function for De Novo Drug Discovery

In the quest for accelerating de novo drug discovery, the development of efficient and accurate scoring functions represents a fundamental challenge. This study introduces iScore, a novel machine learning (ML)-based scoring function designed to predict the binding affinity of protein-ligand complexes with remarkable speed and precision. Uniquely, iScore circumvents the conventional reliance on explicit knowledge of protein-ligand interactions and a full picture of atomic contacts, instead leveraging a set of ligand and binding pocket descriptors to directly evaluate binding affinity. This approach enables skipping the inefficient and slow conformational sampling stage, thereby enabling the rapid screening of ultrahuge molecular libraries, a crucial advancement given the practically infinite dimensions of chemical space. iScore was rigorously trained and validated using the PDBbind 2020 refined set, CASF 2016, CSAR NRC-HiQ Set1/2, DUD-E, and target fishing data sets, employing three distinct ML methodologies: Deep neural network (iScore-DNN), random forest (iScore-RF), and eXtreme gradient boosting (iScore-XGB). A hybrid model, iScore-Hybrid, was subsequently developed to incorporate the strengths of these individual base learners. The hybrid model demonstrated a Pearson correlation coefficient (R) of 0.78 and a root-mean-square error (RMSE) of 1.23 in cross-validation, outperforming the individual base learners and establishing new benchmarks for scoring power (R = 0.814, RMSE = 1.34), ranking power (ρ = 0.705), and screening power (success rate at top 10% = 73.7%). Moreover, iScore-Hybrid demonstrated great performance in the target fishing benchmarking study.

Data-driven approaches for predicting mechanical properties and determining processing parameters of selective laser sintered nylon-12 components

In order to allow engineers to make decisions regarding laser settings in selective laser sintering and predict the mechanical properties of materials, conventional material models could provide accurate solutions and recommendations, however, they are potentially expensive and time-consuming. Thus, a number of computational data-driven methodologies have been introduced in this article, as alternatives, to formulate cross-correlations between the processing parameters and mechanical properties of selective laser sintered (SLS) nylon-12 components. Proposed in this article direct-from laser settings to material properties, and inverse-from desired material properties to laser settings, two estimation frameworks have provided accurate estimation results. The accuracy of three proposed data-driven methodologies: fuzzy inference system (FIS), artificial neural networks (ANN) and adaptive neural fuzzy inference system (ANFIS), have been compared and thoroughly analysed, with FIS being the most accurate solution.

Supplementary Information

The online version contains supplementary material available at 10.1007/s44245-025-00094-7.

A genetic algorithm optimized hybrid model for agricultural price forecasting based on VMD and LSTM network

Accurately predicting agricultural commodity prices is challenging due to their unpredictable and complex nature. Existing models often fail to capture nonlinear and nonstationary patterns in price data, resulting in less accurate forecasts. To tackle these challenges, we present a novel hybrid VMD-LSTM model that synergistically combines genetic algorithm (GA), variational mode decomposition (VMD), and long short-term memory (LSTM), leading to better prediction accuracy. The proposed model utilizes GA-optimized VMD, which decomposes a price series into intrinsic mode functions (IMFs) with a unique property of sparsity leading to faster convergence. Then, these IMFs are individually modelled and forecasted using GA-optimized LSTM models. Finally, the forecasts of all IMFs are ensembled to provide an output for the actual price series. VMD-LSTM is evaluated against individual LSTM and decomposition-based models (EMD-LSTM, EEMD-LSTM, CEEMDAN-LSTM) using monthly price data for maize, palm oil, and soybean oil. Performance is assessed through root mean square error (RMSE), mean absolute percentage error (MAPE), and directional prediction statistics ([Formula: see text]). VMD-LSTM reduces RMSE by 56.93%, 21.83%, and 27.00% and MAPE by 44%, 21.67%, and 25.85% for maize, palm oil, and soybean oil, respectively, compared to the next best CEEMDAN-LSTM. TOPSIS and Diebold-Mariano test also confirm the better prediction accuracy of VMD-LSTM. The proposed model can be a better tool for agricultural price forecasting, supporting decision-making for farmers, traders, and policymakers.

Laser induced forward transfer imaging using deep learning

A novel approach for improving the accuracy and efficiency of laser-induced forward transfer (LIFT), through the application of deep learning techniques is presented. By training a neural network on a dataset of images of donor and receiver substrates, the appearance of copper droplets deposited onto the receiver was predicted directly from images of the donor. The results of droplet image prediction using LIFT gave an average RMSE of 9.63 compared with the experimental images, with the SSIM ranging from 0.75 to 0.83, reflecting reliable structural similarity across predictions. These findings underscore the model's predictive potential while identifying opportunities for refinement in minimising error. This approach has the potential to transform parameter optimisation for LIFT, as it enables the visualization of the deposited material without the time-consuming requirement of removing the donor from the setup to allow inspection of the receiver. This work therefore represents an important step forward in the development of LIFT as an additive manufacturing technology to create complex 3D structures on the microscale.

Transitioning from stress electrocardiogram to cardiopulmonary exercise testing: a paradigm shift toward comprehensive medical evaluation of exercise function

Cardiopulmonary exercise testing (CPET) has emerged as a powerful diagnostic tool, providing comprehensive physiological insights into the integrated function of cardiovascular, respiratory, and metabolic systems. Exploiting physiological interactions, CPET allows in-depth diagnostic insights. CPET performance entrains several complexities. Interpreting CPET data can be challenging, requiring significant physiological expertise. The advent of artificial intelligence (AI) has introduced a transformative approach to CPET interpretation, enhancing accuracy, efficiency, and clinical decision-making. This review article explores the current state of AI applications in CPET, highlighting AI's potential to replace the traditional stress electrocardiogram (ECG) test as the preferred diagnostic tool in preventive medicine and medical screening. The article discusses the underlying principles of AI, its integration into CPET interpretation, and the associated benefits, including improved diagnostic accuracy, reduced interobserver variability, and expedited decision-making. Additionally, it addresses the challenges and considerations surrounding the implementation of AI in CPET such as data quality, model interpretability, and ethical concerns. The review concludes by emphasizing the significant promise of AI-assisted CPET interpretation in revolutionizing preventive medicine and medical screening settings and enhancing patient care.

4D-ONIX for reconstructing 3D movies from sparse X-ray projections via deep learning

The X-ray flux from X-ray free-electron lasers and storage rings enables new spatiotemporal opportunities for studying in-situ and operando dynamics, even with single pulses. X-ray multi-projection imaging is a technique that provides volumetric information using single pulses while avoiding the centrifugal forces induced by conventional time-resolved 3D methods like time-resolved tomography, and can acquire 3D movies (4D) at least three orders of magnitude faster than existing techniques. However, reconstructing 4D information from highly sparse projections remains a challenge for current algorithms. Here we present 4D-ONIX, a deep-learning-based approach that reconstructs 3D movies from an extremely limited number of projections. It combines the computational physical model of X-ray interaction with matter and state-of-the-art deep learning methods. We demonstrate its ability to reconstruct high-quality 4D by generalizing over multiple experiments with only two to three projections per timestamp on simulations of water droplet collisions and experimental data of additive manufacturing. Our results demonstrate 4D-ONIX as an enabling tool for 4D analysis, offering high-quality image reconstruction for fast dynamics three orders of magnitude faster than tomography.

Innovative Applications of Nanopore Technology in Tumor Screening: An Exosome-Centric Approach

Cancer remains one of the leading causes of death worldwide. Its complex pathogenesis and metastasis pose significant challenges for early diagnosis, underscoring the urgent need for innovative and non-invasive tumor screening methods. Exosomes, small extracellular vesicles that reflect the physiological and pathological states of their parent cells, are uniquely suited for cancer liquid biopsy due to their molecular cargo, including RNA, DNA, and proteins. However, traditional methods for exosome isolation and detection are often limited by inadequate sensitivity, specificity, and efficiency. Nanopore technology, characterized by high sensitivity and single-molecule resolution, offers powerful tools for exosome analysis. This review highlights its diverse applications in tumor screening, such as magnetic nanopores for high-throughput sorting, electrochemical sensing for real-time detection, nanomaterial-based assemblies for efficient capture, and plasmon resonance for ultrasensitive analysis. These advancements have enabled precise exosome detection and demonstrated promising potential in the early diagnosis of breast, pancreatic, and prostate cancers, while also supporting personalized treatment strategies. Additionally, this review summarizes commercialized products for exosome-based cancer diagnostics and examines the technical and translational challenges in clinical applications. Finally, it discusses the future prospects of nanopore technology in advancing liquid biopsy toward clinical implementation. The continued progress of nanopore technology not only accelerates exosome-based precision medicine but also represents a significant step forward in next-generation liquid biopsy and tumor screening.

CrowdRadar: a mobile crowdsensing framework for urban traffic green travel safety risk assessment

As environmental awareness increased due to the surge in greenhouse gases, green travel modes such as bicycles and walking have gradually became popular choices. However, the current traffic environment has many hidden problems that endanger the personal safety of traffic participants and hinder the development of green travel. Traditional methods, such as identifying risky locations after traffic accidents, suffer from the disadvantages of delayed response and lack of foresight. Against this background, we proposed a mobile edge crowdsensing framework to dynamically assess urban traffic green travel safety risks. Specifically, a large number of mobile devices were used to sense the road environment, from which a semantic detection framework detected the traffic high-risk behaviors of traffic participants. Then multi-source and heterogeneous urban crowdsensing data were used to model the travel safety risk to achieve a comprehensive and real-time assessment of urban green travel safety. We evaluated our method by leveraging real-world datasets collected from Xiamen Island. Results showed that our framework could accurately detect traffic high-risk behaviors with average F1-scores of 86.5% and assessed the travel safety risk with R 2 of 0.85 outperforming various baseline methods.

Exploring AI in metasurface structures with forward and inverse design

As an artificially manufactured planar device, a metasurface structure can produce unusual electromagnetic responses by harnessing four basic characteristics of the light wave. Traditional design processes rely on numerical algorithms combined with parameter optimization. However, such methods are often time-consuming and struggle to match actual responses. This paper aims to give a unique perspective to classify the artificial intelligence(AI)-enabled design, dividing it into forward and inverse designs according to the mapping relationship between variables and performance. Forward designs are driven by intelligent algorithms; neural networks are one of the principal ways to realize reverse design. This paper reviews recent progress in AI-enabled metasurface design, examining the principles, advantages, and potential applications. A rich content and detailed comparison can help build a holistic understanding of metasurface design. Moreover, the authors believe that this systematic and detailed review will pave the way for future research and the selection of practical applications.

Enhancing malignant transformation predictions in oral potentially malignant disorders: A novel machine learning framework using real-world data

This study addresses the challenge of accurately predicting malignant transformation risk in patients with oral potentially malignant disorders (OPMDs). Using data from 1,094 patients across three institutions (2004-2023), the researchers compared traditional statistical methods, including a Cox proportional hazards (Cox-PH) nomogram, with machine learning (ML) algorithms. A novel Self Attention Artificial Neural Network (SANN) model was developed, trained, and validated alongside other ML models including ANN, RF, and DeepSurv. The SANN model outperformed all other approaches, achieving an AUC of 0.9877, with sensitivity, specificity, accuracy, and precision exceeding 0.96. In comparison, the Cox-PH nomogram achieved AUCs of 0.880-0.902. Comprehensive evaluations using Receiver Operating Characteristic, calibration curves, and decision curve analysis demonstrated SANN's superior predictive efficacy, robustness, and generalizability. These findings highlight the potential of customized ML models, particularly SANN, to enhance early identification and management of high-risk OPMD patients, outperforming conventional statistical methods.

Unveiling CNS cell morphology with deep learning: A gateway to anti-inflammatory compound screening

Deciphering the complex relationships between cellular morphology and phenotypic manifestations is crucial for understanding cell behavior, particularly in the context of neuropathological states. Despite its importance, the application of advanced image analysis methodologies to central nervous system (CNS) cells, including neuronal and glial cells, has been limited. Furthermore, cutting-edge techniques in the field of cell image analysis, such as deep learning (DL), still face challenges, including the requirement for large amounts of labeled data, difficulty in detecting subtle cellular changes, and the presence of batch effects. Our study addresses these shortcomings in the context of neuroinflammation. Using our in-house data and a DL-based approach, we have effectively analyzed the morphological phenotypes of neuronal and microglial cells, both in pathological conditions and following pharmaceutical interventions. This innovative method enhances our understanding of neuroinflammation and streamlines the process for screening potential therapeutic compounds, bridging a gap in neuropathological research and pharmaceutical development.

Machine-learning models for Alzheimer's disease diagnosis using neuroimaging data: survey, reproducibility, and generalizability evaluation

Clinical diagnosis of Alzheimer's disease (AD) is usually made after symptoms such as short-term memory loss are exhibited, which minimizes the intervention and treatment options. The existing screening techniques cannot distinguish between stable MCI (sMCI) cases (i.e., patients who do not convert to AD for at least three years) and progressive MCI (pMCI) cases (i.e., patients who convert to AD in three years or sooner). Delayed diagnosis of AD also disproportionately affects underrepresented and socioeconomically disadvantaged populations. The significant positive impact of an early diagnosis solution for AD across diverse ethno-racial and demographic groups is well-known and recognized. While advancements in high-throughput technologies have enabled the generation of vast amounts of multimodal clinical, and neuroimaging datasets related to AD, most methods utilizing these data sets for diagnostic purposes have not found their way in clinical settings. To better understand the landscape, we surveyed the major preprocessing, data management, traditional machine-learning (ML), and deep learning (DL) techniques used for diagnosing AD using neuroimaging data such as structural magnetic resonance imaging (sMRI), functional magnetic resonance imaging (fMRI), and positron emission tomography (PET). Once we had a good understanding of the methods available, we conducted a study to assess the reproducibility and generalizability of open-source ML models. Our evaluation shows that existing models show reduced generalizability when different cohorts of the data modality are used while controlling other computational factors. The paper concludes with a discussion of major challenges that plague ML models for AD diagnosis and biomarker discovery.

Computational memory capacity predicts aging and cognitive decline

Memory is a crucial cognitive function that deteriorates with age. However, this ability is normally assessed using cognitive tests instead of the architecture of brain networks. Here, we use reservoir computing, a recurrent neural network computing paradigm, to assess the linear memory capacities of neural-network reservoirs extracted from brain anatomical connectivity data in a lifespan cohort of 636 individuals. The computational memory capacity emerges as a robust marker of aging, being associated with resting-state functional activity, white matter integrity, locus coeruleus signal intensity, and cognitive performance. We replicate our findings in an independent cohort of 154 young and 72 old individuals. By linking the computational memory capacity of the brain network with cognition, brain function and integrity, our findings open new pathways to employ reservoir computing to investigate aging and age-related disorders.

Uncovering water conservation patterns in semi-arid regions through hydrological simulation and deep learning

Under the increasing pressure of global climate change, water conservation (WC) in semi-arid regions is experiencing unprecedented levels of stress. WC involves complex, nonlinear interactions among ecosystem components like vegetation, soil structure, and topography, complicating research. This study introduces a novel approach combining InVEST modeling, spatiotemporal transfer of Water Conservation Reserves (WCR), and deep learning to uncover regional WC patterns and driving mechanisms. The InVEST model evaluates Xiong'an New Area's WC characteristics from 2000 to 2020, showing a 74% average increase in WC depth with an inverted "V" spatial distribution. Spatiotemporal analysis identifies temporal changes, spatial patterns of WCR and land use, and key protection areas, revealing that the WCR in Xiong'an New Area primarily shifts from the lowest WCR areas to lower WCR areas. The potential enhancement areas of WCR are concentrated in the northern region. Deep learning quantifies data complexity, highlighting critical factors like land use, precipitation, and drought influencing WC. This detailed approach enables the development of personalized WC zones and strategies, offering new insights into managing complex spatial and temporal WC data.

A Sheep Behavior Recognition Approach Based on Improved FESS-YOLOv8n Neural Network

Sheep are an important breed of livestock in the northern regions of China, providing humans with nutritious meat and by-products. Therefore, it is essential to ensure the health status of sheep. Research has shown that the individual and group behaviors of sheep can reflect their overall health status. However, as the scale of farming expands, traditional behavior detection methods based on manual observation and those that employ contact-based devices face challenges, including poor real-time performance and unstable accuracy, making them difficult to meet the current demands. To address these issues, this paper proposes a sheep behavior detection model, Fess-YOLOv8n, based on an enhanced YOLOv8n neural network. On the one hand, this approach achieves a lightweight model by introducing the FasterNet structure and the selective channel down-sampling module (SCDown). On the other hand, it utilizes the efficient multi-scale attention mechanism (EMA)as well as the spatial and channel synergistic attention module (SCSA) to improve recognition performance. The results on a self-built dataset show that Fess-YOLOv8n reduced the model size by 2.56 MB and increased the detection accuracy by 4.7%. It provides technical support for large-scale sheep behavior detection and lays a foundation for sheep health monitoring.

A survey of model compression techniques: past, present, and future

The exceptional performance of general-purpose large models has driven various industries to focus on developing domain-specific models. However, large models are not only time-consuming and labor-intensive during the training phase but also have very high hardware requirements during the inference phase, such as large memory and high computational power. These requirements pose considerable challenges for the practical deployment of large models. As these challenges intensify, model compression has become a vital research focus to address these limitations. This paper presents a comprehensive review of the evolution of model compression techniques, from their inception to future directions. To meet the urgent demand for efficient deployment, we delve into several compression methods-such as quantization, pruning, low-rank decomposition, and knowledge distillation-emphasizing their fundamental principles, recent advancements, and innovative strategies. By offering insights into the latest developments and their implications for practical applications, this review serves as a valuable technical resource for researchers and practitioners, providing a range of strategies for model deployment and laying the groundwork for future advancements in model compression.

Machine Learning and Artificial Intelligence in the Multi-Omics Approach to Gut Microbiota

The gut microbiome is involved in human health and disease, and its comprehensive understanding is necessary to exploit it as a diagnostic or therapeutic tool. Multi-omics approaches, including metagenomics, metatranscriptomics, metabolomics, and metaproteomics, enable depiction of the gut microbial ecosystem's complexity. However, these tools generate a large data stream in which integration is needed to produce clinically useful readouts, but, in turn, might be difficult to carry out with conventional statistical methods. Artificial intelligence and machine learning have been increasingly applied to multi-omics datasets in several conditions associated with microbiome disruption, from chronic disorders to cancer. Such tools have potential for clinical implementation, including discovery of microbial biomarkers for disease classification or prediction, prediction of response to specific treatments, and fine-tuning of microbiome-modulating therapies. The state of the art, potential, and limits, of artificial intelligence and machine learning in the multi-omics approach to gut microbiome are discussed.

Automated Detection of Hydrocephalus in Pediatric Head Computed Tomography Using VGG 16 CNN Deep Learning Architecture and Based Automated Segmentation Workflow for Ventricular Volume Estimation

Hydrocephalus, particularly congenital hydrocephalus in infants, remains underexplored in deep learning research. While deep learning has been widely applied to medical image analysis, few studies have specifically addressed the automated classification of hydrocephalus. This study proposes a convolutional neural network (CNN) model based on the VGG16 architecture to detect hydrocephalus in infant head CT images. The model integrates an automated method for ventricular volume extraction, applying windowing, histogram equalization, and thresholding techniques to segment the ventricles from surrounding brain structures. Morphological operations refine the segmentation and contours are extracted for visualization and volume measurement. The dataset consists of 105 head CT scans, each with 60 slices covering the ventricular volume, resulting in 6300 slices. Manual segmentation by three trained radiologists served as the reference standard. The automated method showed a high correlation with manual measurements, with R2 values ranging from 0.94 to 0.99. The mean absolute percentage error (MAPE) ranged 3.99 to 11.13%, while the root mean square error (RRMSE) from 4.56 to 13.74%. To improve model robustness, the dataset was preprocessed, normalized, and augmented with rotation, shifting, zooming, and flipping. The VGG16-based CNN used pre-trained convolutional layers with additional fully connected layers for classification, predicting hydrocephalus or normal labels. Performance evaluation using a multi-split strategy (15 independent splits) achieved a mean accuracy of 90.4% ± 1.2%. This study presents an automated approach for ventricular volume extraction and hydrocephalus detection, offering a promising tool for clinical and research applications with high accuracy and reduced observer bias.

Weighted Multi-Modal Contrastive Learning Based Hybrid Network for Alzheimer's Disease Diagnosis

Multiple imaging modalities and specific proteins in the cerebrospinal fluid, providing a comprehensive understanding of neurodegenerative disorders, have been widely used for computer-aided diagnosis of Alzheimer's disease (AD). Given the proven effectiveness of contrastive learning in aligning multi-modal representation, in this paper, we investigate effective contrastive learning strategies to learn better cross-modal representations for the integration of multi-modal complementary information. To enhance the overall performance in AD diagnosis, we construct a unified hybrid network that integrates feature learning and classifier learning into an end-to-end framework. Specifically, we propose a weighted multi-modal contrastive learning based on hybrid network (WMCL-HN) method. Firstly, an adaptive weighted strategy is implemented on the multi-modal contrastive learning to dynamically regulate the degree of information exchange across modalities. It assigns higher weights to more important modality pair, thus the most important underlying relationships across modalities can be captured. Secondly, we construct a hybrid network, which employs a curriculum learning strategy that gradually transitions the training from feature learning to classifier learning, ensuring that the learned features are tailored to the diagnostic task. Experimental results on ADNI dataset demonstrate the effectiveness of the proposed WMCL-HN in AD-related diagnosis tasks. The source code is available at https://github.com/pcehnago/WMCL-HN.

Early-stage fertilised egg viability detection based on machine vision

1. In the early stages of incubation, challenges arise in the intelligent recognition of multiple eggs on the incubation tray and in achieving consistent high-throughput detection. To address these issues, a method was proposed using a monochrome camera to capture transillumination images of eggs. This work examined factors affecting image consistency, such as light source intensity, imaging uniformity and egg positioning and developed a correction algorithm for non-uniform light intensity in the captured images.2. On day 0 of incubation, images of the egg tray and fertilised eggs were acquired. After applying median filtering, Laplacian sharpening and fixed-threshold segmentation, the egg regions from the images were extracted. These regions were then converted into labelled images for circular fitting, with the fitted circles contracted inward by 10 pixels to define the target egg region as the template for viability detection.3. Using these template images, egg regions from days 5 to 9 of incubation were extracted and four greyscale features derived; mean, maximum, minimum and standard deviation, and four texture features; energy, correlation, homogeneity and contrast were used as input parameters for classification models using Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and a custom Convolutional Neural Network (CNN).4. The CNN model demonstrated the best performance, achieving 99% accuracy on day 8, with Precision, Recall and F1 scores of 0.99, 1.00 and 0.99 for viable embryos, respectively. For non-viable and infertile eggs, Precision, Recall and F1 scores were 1.00, 0.95 and 0.98, respectively. The optimal detection time was determined to be day 6, with an accuracy of 95%, which was one day earlier than the optimal manual inspection time.5. These findings showed that using a monochrome camera with image processing and classification models could enable high-throughput, early-stage viability detection of fertilised eggs. This can be used as technical support for the development of automated detection systems.

Memristive neuromorphic interfaces: integrating sensory modalities with artificial neural networks

The advent of the Internet of Things (IoT) has led to exponential growth in data generated from sensors, requiring efficient methods to process complex and unstructured external information. Unlike conventional von Neumann sensory systems with separate data collection and processing units, biological sensory systems integrate sensing, memory, and computing to process environmental information in real time with high efficiency. Memristive neuromorphic sensory systems using memristors as their basic components have emerged as promising alternatives to CMOS-based systems. Memristors can closely replicate the key characteristics of biological receptors, neurons, and synapses by integrating the threshold and adaptation properties of receptors, the action potential firing in neurons, and the synaptic plasticity of synapses. Furthermore, through careful engineering of their switching dynamics, the electrical properties of memristors can be tailored to emulate specific functions, while benefiting from high operational speed, low power consumption, and exceptional scalability. Consequently, their integration with high-performance sensors offers a promising pathway toward realizing fully integrated artificial sensory systems that can efficiently process and respond to diverse environmental stimuli in real time. In this review, we first introduce the fundamental principles of memristive neuromorphic technologies for artificial sensory systems, explaining how each component is structured and what functions it performs. We then discuss how these principles can be applied to replicate the four traditional senses, highlighting the underlying mechanisms and recent advances in mimicking biological sensory functions. Finally, we address the remaining challenges and provide prospects for the continued development of memristor-based artificial sensory systems.

Application of machine learning in the context of reoperation, outcome and management after ACL reconstruction - A systematic review

INTRODUCTION

Machine learning-based tools are becoming increasingly popular in clinical practice. They offer new possibilities but are also limited in their reliability and accuracy.

OBJECTIVES

The present systematic review updates and discusses the existing literature regarding machine learning algorithm-based tools to predict outcome and management in patients after ACL reconstruction.

METHOD

PubMed was searched for articles containing machine learning algorithms related to anterior cruciate ligament reconstruction and its outcome and management. No additional filters or time constraints were used. All eligible studies were accessed by hand.

RESULTS

After screening of 115 articles, 15 were included. Six studies evaluated predictors for reoperation after ACL surgery. Four studies investigated the clinical outcome prediction after ACL reconstruction including prediction of secondary meniscus tear and secondary knee osteoarthritis. Single topics addressed in patients with ACL reconstruction were costs, opioid use, the need for a femoral nerve block, short term complications, hospital admission, and reduction of the burden to complete the Knee Osteoarthritis and Outcome score questionnaire. Predictive power was very heterogeneous, depending on the specific research question and parameters included.

CONCLUSION

New machine-learning tools offer interesting insights into variables affecting the target outcome and general management of patients with ACL reconstruction. While present machine-learning based prediction models seem to outperform previous existing benchmark regression models, their predictive ability often is still too low to base individual decision making on them. With the rapid progress observed over the last few years, it is conceivable that this might change, however, in the foreseeable future.

Emerging Liquid-Based Memristive Devices for Neuromorphic Computation

To mimic the neural functions of the human brain, developing hardware with natural similarities to the human nervous system is crucial for realizing neuromorphic computing architectures. Owing to their capability to emulate artificial neurons and synapses, memristors are widely regarded as a leading candidate for achieving neuromorphic computing. However, most current memristor devices are solid-state. In contrast, biological nervous systems operate within an aqueous environment, and the human brain accomplishes intelligent behaviors such as information generation, transmission, and memory by regulating ion transport in neuronal cells. To achieve computing systems that are more analogous to biological systems and more energy-efficient, memristor devices based on liquid environments are developed. In contrast to traditional solid-state memristors, liquid-based memristors possess advantages such as anti-interference, low energy consumption, and low heat generation. Simultaneously, they demonstrate excellent biocompatibility, rendering them an ideal option for the next generation of artificial intelligence systems. Numerous experimental demonstrations of liquid-based memristors are reported, showcasing their unique memristive properties and novel neuromorphic functionalities. This review focuses on the recent developments in liquid-based memristors, discussing their operating mechanisms, structures, and functional characteristics. Additionally, the potential applications and development directions of liquid-based memristors in neuromorphic computing systems are proposed.

Optimizing deep neural networks for high-resolution land cover classification through data augmentation

This study presents an innovative approach to high-resolution land cover classification using deep learning, tackling the challenge of working with an exceptionally small dataset. Manual annotation of land cover data is both time-consuming and labor-intensive, making data augmentation crucial for enhancing model performance. While data augmentation is a well-established technique, there has not been a comprehensive and comparative evaluation of a wide range of data augmentation methods specifically applied to land cover classification until now. Our work fills this gap by systematically testing eight different data augmentation techniques across four neural networks (U-Net, DeepLabv3 + , FCN, PSPNet) using 25 cm resolution images from Cantabria, Spain. In total, we generated 19 distinct training sets and trained and validated 72 models. The results show that data augmentation can boost model performance by up to 30%. The best model (DeepLabV3 + with flip, contrast, and brightness adjustments) achieved an accuracy of 0.89 and an IoU of 0.78. Additionally, we utilized this optimized model to generate land cover maps for the years 2014, 2017, and 2019, validated at 580 samples selected based on a stratified sampling approach using CORINE Land Cover data, achieving an accuracy of 87.2%. This study not only provides a systematic ranking of data augmentation techniques for land cover classification but also offers a practical framework to help future researchers save time by identifying the most effective augmentation strategies for this specific task.

Nearly Optimal Learning Using Sparse Deep ReLU Networks in Regularized Empirical Risk Minimization With Lipschitz Loss

We propose a sparse deep ReLU network (SDRN) estimator of the regression function obtained from regularized empirical risk minimization with a Lipschitz loss function. Our framework can be applied to a variety of regression and classification problems. We establish novel nonasymptotic excess risk bounds for our SDRN estimator when the regression function belongs to a Sobolev space with mixed derivatives. We obtain a new, nearly optimal, risk rate in the sense that the SDRN estimator can achieve nearly the same optimal minimax convergence rate as one-dimensional nonparametric regression with the dimension involved in a logarithm term only when the feature dimension is fixed. The estimator has a slightly slower rate when the dimension grows with the sample size. We show that the depth of the SDRN estimator grows with the sample size in logarithmic order, and the total number of nodes and weights grows in polynomial order of the sample size to have the nearly optimal risk rate. The proposed SDRN can go deeper with fewer parameters to well estimate the regression and overcome the overfitting problem encountered by conventional feedforward neural networks.

Deep Learning-Based Semantic Segmentation for Objective Colonoscopy Quality Assessment

Background: This study aims to objectively evaluate the overall quality of colonoscopies using a specially trained deep learning-based semantic segmentation neural network. This represents a modern and valuable approach for the analysis of colonoscopy frames. Methods: We collected thousands of colonoscopy frames extracted from a set of video colonoscopy files. A color-based image processing method was used to extract color features from specific regions of each colonoscopy frame, namely, the intestinal mucosa, residues, artifacts, and lumen. With these features, we automatically annotated all the colonoscopy frames and then selected the best of them to train a semantic segmentation network. This trained network was used to classify the four region types in a different set of test colonoscopy frames and extract pixel statistics that are relevant to quality evaluation. The test colonoscopies were also evaluated by colonoscopy experts using the Boston scale. Results: The deep learning semantic segmentation method obtained good results, in terms of classifying the four key regions in colonoscopy frames, and produced pixel statistics that are efficient in terms of objective quality assessment. The Spearman correlation results were as follows: BBPS vs. pixel scores: 0.69; BBPS vs. mucosa pixel percentage: 0.63; BBPS vs. residue pixel percentage: -0.47; BBPS vs. Artifact Pixel Percentage: -0.65. The agreement analysis using Cohen's Kappa yielded a value of 0.28. The colonoscopy evaluation based on the extracted pixel statistics showed a fair level of compatibility with the experts' evaluations. Conclusions: Our proposed deep learning semantic segmentation approach is shown to be a promising tool for evaluating the overall quality of colonoscopies and goes beyond the Boston Bowel Preparation Scale in terms of assessing colonoscopy quality. In particular, while the Boston scale focuses solely on the amount of residual content, our method can identify and quantify the percentage of colonic mucosa, residues, and artifacts, providing a more comprehensive and objective evaluation.

8-bit states in 2D floating-gate memories using gate-injection mode for large-scale convolutional neural networks

The fast development of artificial intelligence has called for high-efficiency neuromorphic computing hardware. While two-dimensional floating-gate memories show promise, their limited state numbers and stability hinder practical use. Here, we report gate-injection-mode two-dimensional floating-gate memories as a candidate for large-scale neural network accelerators. Through a coplanar device structure design and a bi-pulse state programming strategy, 8-bit states with intervals larger than three times of the standard deviations and stability over 10,000 s are achieved at 3 V. The cycling endurance is over 105 and the fabricated 256 devices show a yield of 94.9%. Leveraging this, we carry out experimental image convolutions and 38,592 kernels transplanting on an integrated 9 × 2 array that exhibits results matching well with simulations. We also show that fix-point neural networks with 8-bit precision have inference accuracies approaching the ideal values. Our work validates the potential of gate-injection-mode two-dimensional floating-gate memories for high-efficiency neuromorphic computing hardware.

Deep image features sensing with multilevel fusion for complex convolution neural networks & cross domain benchmarks

Efficient image retrieval from a variety of datasets is crucial in today's digital world. Visual properties are represented using primitive image signatures in Content Based Image Retrieval (CBIR). Feature vectors are employed to classify images into predefined categories. This research presents a unique feature identification technique based on suppression to locate interest points by computing productive sum of pixel derivatives by computing the differentials for corner scores. Scale space interpolation is applied to define interest points by combining color features from spatially ordered L2 normalized coefficients with shape and object information. Object based feature vectors are formed using high variance coefficients to reduce the complexity and are converted into bag-of-visual-words (BoVW) for effective retrieval and ranking. The presented method encompass feature vectors for information synthesis and improves the discriminating strength of the retrieval system by extracting deep image features including primitive, spatial, and overlayed using multilayer fusion of Convolutional Neural Networks(CNNs). Extensive experimentation is performed on standard image datasets benchmarks, including ALOT, Cifar-10, Corel-10k, Tropical Fruits, and Zubud. These datasets cover wide range of categories including shape, color, texture, spatial, and complicated objects. Experimental results demonstrate considerable improvements in precision and recall rates, average retrieval precision and recall, and mean average precision and recall rates across various image semantic groups within versatile datasets. The integration of traditional feature extraction methods fusion with multilevel CNN advances image sensing and retrieval systems, promising more accurate and efficient image retrieval solutions.

Deep Learning Enhanced Near Infrared-II Imaging and Image-Guided Small Interfering Ribonucleic Acid Therapy of Ischemic Stroke

Small interfering RNA (siRNA) targeting the NOD-like receptor family pyrin domain-containing 3 (NLRP3) inflammasome has emerged as a promising therapeutic strategy to mitigate infarct volume and brain injury following ischemic stroke. However, the clinical translation of siRNA-based therapies is significantly hampered by the formidable blood-brain barrier (BBB), which restricts drug penetration into the central nervous system. To address this challenge, we have developed an innovative long-circulating near-infrared II (NIR-II) nanoparticle platform YWFC NPs, which is meticulously engineered to enhance BBB transcytosis and enable efficient delivery of siRNA targeting NLRP3 (siNLRP3@YWFC NPs) in preclinical models of ischemic stroke. Furthermore, we integrated advanced deep learning neural network algorithms to optimize in vivo NIR-II imaging of the cerebral infarct penumbra, achieving an improved signal-to-background ratio at 72 h poststroke. In vivo studies employing middle cerebral artery occlusion (MCAO) mouse models demonstrated that image-guided therapy with siNLRP3@YWFC NPs, guided by prolonged NIR-II imaging, resulted in significant therapeutic benefits.

Assessments of lung nodules by an artificial intelligence chatbot using longitudinal CT images

Large language models have shown efficacy across multiple medical tasks. However, their value in the assessment of longitudinal follow-up computed tomography (CT) images of patients with lung nodules is unclear. In this study, we evaluate the ability of the latest generative pre-trained transformer (GPT)-4o model to assess changes in malignancy probability, size, and features of lung nodules on longitudinal CT scans from 647 patients (547 from two local centers and 100 from a public dataset). GPT-4o achieves an average accuracy of 0.88 in predicting lung nodule malignancy compared to pathological results and an average intraclass correlation coefficient of 0.91 in measuring nodule size compared with manual measurements by radiologists. Six radiologists' evaluations demonstrate GPT-4o's ability to capture changes in nodule features with a median Likert score of 4.17 (out of 5.00). In summary, GPT-4o could capture dynamic changes in lung nodules across longitudinal follow-up CT images, thus providing high-quality radiological evidence to assist in clinical management.

Enhancing Mild Cognitive Impairment Auxiliary Identification Through Multimodal Cognitive Assessment with Eye Tracking and Convolutional Neural Network Analysis

Background: Mild Cognitive Impairment (MCI) is a critical transitional phase between normal aging and dementia, and early detection is essential to mitigate cognitive decline. Traditional cognitive assessment tools, such as the Mini-Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCA), exhibit limitations in feasibility, which potentially and partially affects results for early-stage MCI detection. This study developed and tested a supportive cognitive assessment system for MCI auxiliary identification, leveraging eye-tracking features and convolutional neural network (CNN) analysis. Methods: The system employed eye-tracking technology in conjunction with machine learning to build a multimodal auxiliary identification model. Four eye movement tasks and two cognitive tests were administered to 128 participants (40 MCI patients, 57 elderly controls, 31 young adults as reference). We extracted 31 eye movement and 8 behavioral features to assess their contributions to classification accuracy using CNN analysis. Eye movement features only, behavioral features only, and combined features models were developed and tested respectively, to find out the most effective approach for MCI auxiliary identification. Results: Overall, the combined features model achieved a higher discrimination accuracy than models with single feature sets alone. Specifically, the model's ability to differentiate MCI from healthy individuals, including young adults, reached an average accuracy of 74.62%. For distinguishing MCI from elderly controls, the model's accuracy averaged 66.50%. Conclusions: Results show that a multimodal model significantly outperforms single-feature models in identifying MCI, highlighting the potential of eye-tracking for early detection. These findings suggest that integrating multimodal data can enhance the effectiveness of MCI auxiliary identification, providing a novel potential pathway for community-based early detection efforts.

The impact of multi-modality fusion and deep learning on adult age estimation based on bone mineral density

INTRODUCTION

Age estimation, especially in adults, presents substantial challenges in different contexts ranging from forensic to clinical applications. Bone mineral density (BMD), with its distinct age-related variations, has emerged as a critical marker in this domain. This study aims to enhance chronological age estimation accuracy using deep learning (DL) incorporating a multi-modality fusion strategy based on BMD.

METHODS

We conducted a retrospective analysis of 4296 CT scans from a Chinese population, covering August 2015 to November 2022, encompassing lumbar, femur, and pubis modalities. Our DL approach, integrating multi-modality fusion, was applied to predict chronological age automatically. The model's performance was evaluated using an internal real-world clinical cohort of 644 scans (December 2022 to May 2023) and an external cadaver validation cohort of 351 scans.

RESULTS

In single-modality assessments, the lumbar modality excelled. However, multi-modality models demonstrated superior performance, evidenced by lower mean absolute errors (MAEs) and higher Pearson's R² values. The optimal multi-modality model exhibited outstanding R² values of 0.89 overall, 0.88 in females, 0.90 in males, with the MAEs of 4.05 overall, 3.69 in females, 4.33 in males in the internal validation cohort. In the external cadaver validation, the model maintained favourable R² values (0.84 overall, 0.89 in females, 0.82 in males) and MAEs (5.01 overall, 4.71 in females, 5.09 in males), highlighting its generalizability across diverse scenarios.

CONCLUSION

The integration of multi-modalities fusion with DL significantly refines the accuracy of adult age estimation based on BMD. The AI-based system that effectively combines multi-modalities BMD data, presenting a robust and innovative tool for accurate AAE, poised to significantly improve both geriatric diagnostics and forensic investigations.

Histogram matching-enhanced adversarial learning for unsupervised domain adaptation in medical image segmentation

BACKGROUND

Unsupervised domain adaptation (UDA) seeks to mitigate the performance degradation of deep neural networks when applied to new, unlabeled domains by leveraging knowledge from source domains. In medical image segmentation, prevailing UDA techniques often utilize adversarial learning to address domain shifts for cross-modality adaptation. Current research on adversarial learning tends to adopt increasingly complex models and loss functions, making the training process highly intricate and less stable/robust. Furthermore, most methods primarily focused on segmentation accuracy while neglecting the associated confidence levels and uncertainties.

PURPOSE

To develop a simple yet effective UDA method based on histogram matching-enhanced adversarial learning (HMeAL-UDA), and provide comprehensive uncertainty estimations of the model predictions.

METHODS

Aiming to bridge the domain gap while reducing the model complexity, we developed a novel adversarial learning approach to align multi-modality features. The method, termed HMeAL-UDA, integrates a plug-and-play histogram matching strategy to mitigate domain-specific image style biases across modalities. We employed adversarial learning to constrain the model in the prediction space, enabling it to focus on domain-invariant features during segmentation. Moreover, we quantified the model's prediction confidence using Monte Carlo (MC) dropouts to assess two voxel-level uncertainty estimates of the segmentation results, which were subsequently aggregated into a volume-level uncertainty score, providing an overall measure of the model's reliability. The proposed method was evaluated on three public datasets (Combined Healthy Abdominal Organ Segmentation [CHAOS], Beyond the Cranial Vault [BTCV], and Abdominal Multi-Organ Segmentation Challenge [AMOS]) and one in-house clinical dataset (UTSW). We used 30 MRI scans (20 from the CHAOS dataset and 10 from the in-house dataset) and 30 CT scans from the BTCV dataset for UDA-based, cross-modality liver segmentation. Additionally, 240 CT scans and 60 MRI scans from the AMOS dataset were utilized for cross-modality multi-organ segmentation. The training and testing sets for each modality were split with ratios of approximately 4:1-3:1.

RESULTS

Extensive experiments on cross-modality medical image segmentation demonstrated the superiority of HMeAL-UDA over two state-of-the-art approaches. HMeAL-UDA achieved a mean (± s.d.) Dice similarity coefficient (DSC) of 91.34% ± 1.23% and an HD95 of 6.18 ± 2.93 mm for cross-modality (from CT to MRI) adaptation of abdominal multi-organ segmentation, and a DSC of 87.13% ± 3.67% with an HD95 of 2.48 ± 1.56 mm for segmentation adaptation in the opposite direction (MRI to CT). The results are approaching or even outperforming those of supervised methods trained with "ground-truth" labels in the target domain. In addition, we provide a comprehensive assessment of the model's uncertainty, which can help with the understanding of segmentation reliability to guide clinical decisions.

CONCLUSION

HMeAL-UDA provides a powerful segmentation tool to address cross-modality domain shifts, with the potential to generalize to other deep learning applications in medical imaging.

Cauchy activation function and XNet

We have developed a novel activation function, named the Cauchy Activation Function. This function is derived from the Cauchy Integral Theorem in complex analysis and is specifically tailored for problems requiring high precision. This innovation has led to the creation of a new class of neural networks, which we call (Comple)XNet, or simply XNet. We will demonstrate that XNet is particularly effective for high-dimensional challenges such as image classification and solving Partial Differential Equations (PDEs). Our evaluations show that XNet significantly outperforms established benchmarks like MNIST and CIFAR-10 in computer vision, and offers substantial advantages over Physics-Informed Neural Networks (PINNs) in both low-dimensional and high-dimensional PDE scenarios.

Development of a Disease Model for Predicting Postoperative Delirium Using Combined Blood Biomarkers

OBJECTIVE

Postoperative delirium, a common neurocognitive complication after surgery and anesthesia, requires early detection for potential intervention. Herein, we constructed a multidimensional postoperative delirium risk-prediction model incorporating multiple demographic parameters and blood biomarkers to enhance prediction accuracy.

METHODS

We included 555 patients undergoing radical surgery for colorectal cancer. Demographic characteristics and lipid profiles were collected preoperatively, and perioperative anesthesia and surgical conditions were recorded; blood biomarkers were measured before and after surgery. The 3D-CAM scale was used to assess postoperative delirium occurrence within 3 days after surgery. Patients were divided into the postoperative delirium (N = 100) and non-postoperative delirium (N = 455) groups. Based on machine learning, linear and nine non-linear models were developed and compared to select the optimal model. Shapley value-interpretation methods and mediation analysis were used to assess feature importance and interaction.

RESULTS

The median age of the participants was 65 years (interquartile range: 56-71 years; 57.8% male). Among the 10 machine-learning models, the random forest model performed the best (validation cohort, area under the receiver operating characteristic curve of 0.795 [0.704-0.885]). Lipid profile (total cholesterol, triglycerides, and trimethylamine-N-oxide) levels were identified as key postoperative delirium predictors. Mediation analysis further confirmed mediating effects among total cholesterol, trimethylamine-N-oxide, and postoperative delirium; a nomogram model was developed as a web-based tool for external validation and use by other clinicians.

INTERPRETATION

Blood biomarkers are crucial in predicting postoperative delirium and aid anesthesiologists in identifying its risks in a timely manner. This model facilitates personalized perioperative management and reduces the occurrence of postoperative delirium.

TRIAL REGISTRATION

ChiCTR2300075723.

A clinical data-driven machine learning approach for predicting the effectiveness of piperacillin-tazobactam in treating lower respiratory tract infections

BACKGROUND

In hospitalized patients, inadequate antibiotic dosage leading to bacterial resistance and increased antimicrobial use intensity due to overexposure to antibiotics are common problems. In the present study, we constructed a machine learning model based on patients' clinical information to predict the clinical effectiveness of Piperacillin-tazobactam (TZP) (4:1) in treating bacterial lower respiratory tract infections (LRTIs), to assist clinicians in making better clinical decisions.

METHODS

We collected data from patients diagnosed with LRTIs or equivalent diagnoses admitted to the Department of Pulmonary and Critical Care Medicine at Shanghai Pudong Hospital, Shanghai, between January 1, 2021, and July 31, 2023. A total of 26 relevant clinical features were extracted from this cohort. Following data preprocessing, we trained four models: Logistic Regression, Random Forest, Support Vector Machine, and Gaussian Naive Bayes. The dataset was split into training and test sets using a 7:3 ratio. The top-performing models, as determined by Receiver Operating Characteristic (ROC)-Area Under the Curve (AUC) on the independent test set, were subsequently ensembled. Ensemble model (EL) performance was evaluated using bootstrap resampling on the training set and ROC-AUC, recall, accuracy, precision, F1-score, and log loss on an independent test set. The optimal model was then deployed as a web application for clinical outcome prediction.

RESULTS

A total of 1,314 patients primarily treated with TZP as initial empiric antibiotic therapy were enrolled in the analysis. The success group comprised 995 patients (75.7%), while the failure group consisted of 319 patients (24.3%). We constructed an ensemble learning model based on the Logistic Regression, Support Vector Machine and Random Forest models, which showed better overall performance. The EL model demonstrated robust performance on an independent test set, exhibiting a ROC-AUC of 0.69, a recall of 0.69, an accuracy of 0.64, a precision of 0.40, a F1-score of 0.50, and a log loss of 0.66. A corresponding web application was then developed and made available at http://106.12.146.54:1020/ .

CONCLUSIONS

In this study, we successfully developed and validated an EL model that effectively predicts the clinical effectiveness of TZP (4:1) in treating bacterial LRTIs. The model achieved a balanced performance across key evaluation metrics, demonstrating the model's potential utility in clinical decision-making. The web-based application makes this model readily accessible to clinicians, potentially helping optimize antibiotic dosing decisions and reduce both inadequate treatment and overexposure. While promising, future studies with larger datasets and prospective validation are needed to further improve the model's performance and validate its clinical utility. This work represents a step forward in using machine learning to support antimicrobial stewardship and personalized antibiotic therapy.

Establishment and validation of a ResNet-based radiomics model for predicting prognosis in cervical spinal cord injury patients

Cervical spinal cord injury (cSCI) poses a significant challenge due to the unpredictable nature of recovery, which ranges from mild paralysis to severe long-term disability. Accurate prognostic models are crucial for guiding treatment and rehabilitation but are often limited by their reliance on clinical observations alone. Recent advancements in radiomics and deep learning have shown promise in enhancing prognostic accuracy by leveraging detailed imaging data. However, integrating these imaging features with clinical data remains an underexplored area. This study aims to develop a combined model using imaging and clinical signatures to predict the prognosis of cSCI patients six months post-injury, helping clinical decisions and improving rehabilitation plans. We retrospectively analyzed 168 cSCI patients treated at Zhongda Hospital from January 1, 2018, to June 30, 2023. The retrospective cohort was divided into training (134 patients) and testing sets (34 patients) to construct the model. An additional prospective cohort of 43 cSCI patients treated from July 1, 2023, to November 30, 2023, was used as a validation set. Radiomics features were extracted using Pyradiomics and ResNet deep learning from MR images. Clinical factors such as age, smoking history, drinking history, hypertension, diabetes, cardiovascular disease, traumatic brain injury, injury site, and treatment type were analyzed. The LASSO algorithm selected features for model building. Multiple machine learning models, including SVM, LR, NaiveBayes, KNN, RF, ExtraTrees, XGBoost, LightGBM, GradientBoosting, AdaBoosting, and MLP, were used. Receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA) assessed the models' performance. A nomogram was created to visualize the combined model. In Radiomics models, the SVM classifier achieved the highest area under the curve (AUC) of 1.000 in the training set and 0.915 in the testing set. Age, diabetes, and treatment were found clinical risk factors to develop a clinical model. The combined model, integrating radiomics and clinical features, showed strong performance with AUCs of 1.000 in the training set, 0.952 in the testing set and 0.815 in the validation set. And calibration curves and DCA confirmed the model's accuracy and clinical usefulness. This study shows the potential of a combined radiomics and clinical model to predict the prognosis of cSCI patients.