Cesk Patol. 2025;61(2):70-90.

ABSTRACT

The study focuses on the utilization of artificial intelligence (AI) algorithms in the diagnosis of breast, lung, and prostate cancer. It describes the historical development of the digitalization of pathological processes, the implementation of artificial intelligence, and its current applications in pathology. The study emphasizes machine learning, deep learning, computer vision, and digital pathology, which contribute to the automation and refinement of diagnostics. Special attention is given to specific tools such as the uPath systems from Roche and IBEX Medical Analytics, which enable the analysis of histopathological images, tumor cell classification, and biomarker evaluation. The study also highlights the benefits of AI utilization, including increased diagnostic accuracy and efficiency in laboratory processes, while simultaneously addressing the challenges associated with its implementation, such as ethical and legal considerations, data protection, and liability for errors. The aim of this study is to provide a comprehensive overview of the potential applications of AI in digital pathology and its role in modern oncological diagnostics.

PMID:40763009

J Imaging Inform Med. 2025 Aug 5. doi: 10.1007/s10278-025-01632-1. Online ahead of print.

ABSTRACT

Adequate withdrawal time is crucial in colonoscopy, as it is directly associated with polyp detection rates. However, traditional withdrawal time measurements can be biased by non-observation activities, leading to inaccurate assessments of procedural quality. This study aimed to develop a deep learning (DL) model that accurately measures net withdrawal time by excluding non-observation phases and generates quantitative visual summaries of key procedural events. We developed a DL-based automated temporal video segmentation model trained on 40 full-length colonoscopy videos and 825 cecum clips extracted from 221 colonoscopy procedures. The model classifies four key events: cecum, intervention, outside, and narrow-band imaging (NBI) mode. Using the temporal video segmentation results, we calculated the net withdrawal time and extracted representative images from each segment for video summarization. Model performance was evaluated using four standard temporal video segmentation metrics, and its correlation with endoscopist-recorded times on both internal and external test datasets. In both internal and external tests, the DL model achieved a total F1 score exceeding 93% for temporal video segmentation performance. The net withdrawal time showed a strong correlation with endoscopist-recorded times (internal dataset, r = 0.984, p < 0.000; external dataset, r = 0.971, p < 0.000). Additionally, the model successfully generated representative images, and the endoscopists' visual assessment confirmed that these images provided accurate summaries of key events. Compared to manual review, the proposed model offers a more efficient, standardized and objective approach to assessing procedural quality. This model has the potential to enhance clinical practice and improve quality assurance in colonoscopy.

PMID:40762931 | DOI:10.1007/s10278-025-01632-1

Mol Inform. 2025 Aug;44(8):e202500094. doi: 10.1002/minf.70005.

ABSTRACT

In computer-aided drug design, molecular generation models play a crucial role in accelerating the drug development process. Current models mainly fall into two categories: deep learning models with high performance but poor interpretability and heuristic algorithms with better interpretability but limited performance. In this study, we introduce an innovative molecular generation model, the compound construction model (CCMol), which integrates the powerful generative capabilities of the generative pretrained transformer (GPT) and the efficient optimization mechanisms of genetic algorithms (GA) to achieve effective and innovative molecular structures. Specifically, our approach uses structure-based drug design comprising both ligand and protein primary structure-based aspects. CCMol integrates GPT for initial molecular generation and GA for iterative optimization of physicochemical and biological properties. The model's reliability was validated by generating molecules targeting three critical disease-related proteins (GLP1, WRN, and JAK2). The results indicate that CCMol is on average with current advanced models in multiple indicators and performs better than the baseline model in terms of structure diversity and drug-related properties indicators, demonstrating that CCMol exhibits outstanding performance in developing novel and effective candidate drug molecules, particularly suitable for expanding the chemical validity of candidate structures at the early stages of drug discovery.

PMID:40762910 | DOI:10.1002/minf.70005

Radiology. 2025 Aug;316(2):e243393. doi: 10.1148/radiol.243393.

ABSTRACT

Background Sybil, an open-source deep learning model that uses low-dose CT (LDCT) for lung cancer prediction, requires rigorous external testing to confirm generalizability. Additionally, its utility in identifying individuals with high risk who never smoked or have light smoking histories remains unanswered. Purpose To externally test Sybil for identifying individuals with high risk for lung cancer within an Asian health checkup cohort. Materials and Methods This retrospective study analyzed LDCT scans from a single medical checkup facility in a study sample of individuals aged 50-80 years, collected between January 2004 and December 2021, with at least one follow-up scan. The predictive performance of the model for lung cancer risk over a 6-year period was assessed using the time-dependent area under the receiver operating characteristic curve (AUC). These evaluations were conducted in the overall study sample and within subgroups of patients with heavy (at least 20 pack-years) and never- or light smoking histories (ie, ever smoking [median, 2 pack-years]; ineligible for lung cancer screening per 2021 U.S. Preventive Services Task Force recommendations). Additionally, performance was evaluated according to the visibility of lung cancers on baseline LDCT scans. Results: Among 18 057 individuals (median age, 56 years [IQR, 52-61 years]; 11 267 male), 92 lung cancers were diagnosed (0.5%) within 6 years. Of these, 2848 had heavy smoking histories and 9943 had never- or light smoking histories, with 24 (0.8%) and 41 (0.4%) lung cancers, respectively. Sybil achieved AUCs of 0.91 for 1-year risk and 0.74 for 6-year risk. In the heavy-smoking subgroup, 1-year AUC was 0.94 (for visible lung cancers) and 6-year AUC was 0.70 (for future lung cancers). For the never- or light-smoking subgroup, Sybil had an AUC of 0.89 for visible lung cancers and 0.56 for future lung cancers. Conclusion: Sybil demonstrated excellent discriminative performance for visible lung cancers and acceptable performance for future lung cancers in Asian individuals with heavy smoking history but demonstrated poor performance for future lung cancers in a never- or light-smoking subgroup. © RSNA, 2025 Supplemental material is available for this article. See also the editorial by Jacobson and Byrne in this issue.

PMID:40762850 | DOI:10.1148/radiol.243393

Radiology. 2025 Aug;316(2):e243412. doi: 10.1148/radiol.243412.

ABSTRACT

Background Artificial intelligence may enhance diagnostic accuracy in classifying ovarian lesions on MRI scans; however, its applicability across diverse datasets is uncertain. Purpose To develop an efficient, generalizable pipeline for MRI-based ovarian lesion characterization. Materials and Methods In this retrospective study, multiparametric MRI datasets of patients with ovarian lesions from a primary institution (January 2008 to January 2019) and two external institutions (January 2010 to October 2020) were analyzed. Lesions were automatically segmented using Meta's Segment Anything Model (SAM). A DenseNet-121 deep learning (DL) model incorporating both imaging and clinical data was then trained and validated externally for ovarian lesion classification. Lesions were evaluated by radiologists using the Ovarian-Adnexal Reporting and Data System for MRI and subjective assessment, classifying them as benign or malignant. The classification performances of the DL model and radiologists were compared using the DeLong test. Results The primary dataset included 534 lesions from 448 women (mean age, 52 years ± 15 [SD]) from institution A (United States), whereas the external datasets included 58 lesions from 55 women (mean age, 51 years ± 19) from institution B (United States) and 29 lesions from 29 women (mean age, 49 years ± 10) from institution C (Taiwan). SAM-assisted segmentation had a Dice coefficient of 0.86-0.88, reducing the processing time per lesion by 4 minutes compared with manual segmentation. The DL classification model achieved an area under the receiver operating characteristic curve (AUC) of 0.85 (95% CI: 0.85, 0.85) on the internal test and 0.79 (95% CI: 0.79, 0.79 and 0.78, 0.79) across both external datasets with SAM-segmented images, comparable with the radiologists' performance (AUC: 0.84-0.93; all P > .05). Conclusion These results describe an accurate, efficient pipeline that integrates SAM with DL-based classification for differentiating malignant from benign ovarian lesions on MRI scans. It reduced segmentation time and achieved classification performance comparable with that of radiologists. © RSNA, 2025 Supplemental material is available for this article. See also the editorial by Bhayana and Wang in this issue.

PMID:40762846 | DOI:10.1148/radiol.243412

Int Ophthalmol. 2025 Aug 5;45(1):324. doi: 10.1007/s10792-025-03660-w.

ABSTRACT

OBJECTIVE: Ophthalmologists use retinal fundus imaging as a useful tool to diagnose retinal issues. Recently, research on machine learning has concentrated on disease diagnosis. However, disease detection is less accurate, more likely to be misidentified, and often takes a long time to get the right conclusions. This study suggested a new hybrid Deep Learning (DL) approach for retinal illness classification using retinal images to overcome these problems. Three crucial stages are included in this proposed study: preprocessing, feature extraction, and disease classification.

METHODS: At first, the retinal images are preprocessed using the Modified Gaussian Filtering technique to enhance the quality of the image. Subsequently, ResNet, VGG16-based feature descriptors are applied to the preprocessed image along with Improved Multi-Texton features, and statistical features are derived to obtain the most pertinent characteristics and minimize the dimensionality to boost the performance of the model. Then, these obtained features are employed in the hybrid classification model, which is a combination of an Improved LinkNet (ILinkNet) and SqueezeNet models. These models independently process the features for effective classification of disease. Lastly, the final classification results are obtained by averaging the outcomes of both classifiers.

RESULTS: Additionally, the efficiency of the proposed ILink-SqNet model is assessed in comparison to the current techniques. As a result, the ILink-SqNet model achieved a precision of 0.951, which surpasses the result of MobileNet (0.846), SpinalNet (0.821), CNN-Trans (0.836), and LinkNet (0.859), SqueezeNet (0.794) and Fundus-DeepNet (0.762) respectively.

CONCLUSION: Therefore, the suggested ILink-SqNet method provides a robust and effective solution for disease classification, ultimately contributing to better patient outcomes and more efficient clinical practices.

PMID:40762730 | DOI:10.1007/s10792-025-03660-w

Clin Oral Investig. 2025 Aug 5;29(8):396. doi: 10.1007/s00784-025-06485-0.

ABSTRACT

OBJECTIVES: Precisely diagnosing skeletal class is mandatory for correct orthodontic treatment. Artificial intelligence (AI) could increase efficiency during diagnostics and contribute to automated workflows. So far, no AI-driven process can differentiate between skeletal classes I, II, and III in German orthodontic patients. This prospective cross-sectional study aimed to develop machine- and deep-learning models for diagnosing their skeletal class based on the gold-standard individualised ANB of Panagiotidis and Witt.

MATERIALS AND METHODS: Orthodontic patients treated in Germany contributed to the study population. Pre-treatment cephalometric parameters, sex, and age served as input variables. Machine-learning models performed were linear discriminant analysis (LDA), random forest (RF), decision tree (DT), K-nearest neighbours (KNN), support vector machine (SVM), Gaussian naïve Bayes (NB), and multi class logistic regression (MCLR). Furthermore, an artificial neural network (ANN) was conducted.

RESULTS: 1277 German patients presented skeletal class I (48.79%), II (27.56%) and III (23.64%). The best machine-learning model, which considered all input parameters, was RF with 100% accuracy, with Calculated_ANB being the most important (0.429). The model with Calculated_ANB only achieved 100% accuracy (KNN), but ANB alone was inappropriate (71-76% accuracy). The ANN with all parameters and Calculated_ANB achieved 95.31% and 100% validation-accuracy, respectively.

CONCLUSIONS: Machine- and deep-learning methods can correctly determine an individual's skeletal class. Calculated_ANB was the most important among all input parameters, which, therefore, requires precise determination.

CLINICAL RELEVANCE: The AI methods introduced may help to establish digital and automated workflows in cephalometric diagnostics, which could contribute to the relief of the orthodontic practitioner.

PMID:40762676 | DOI:10.1007/s00784-025-06485-0

Anal Chem. 2025 Aug 5. doi: 10.1021/acs.analchem.5c02524. Online ahead of print.

ABSTRACT

Lung cancer, a leading cause of global cancer-related mortality, predominantly features nonsmall cell lung cancer (NSCLC), constituting 80% of all lung malignancies. Despite chemotherapy being the primary NSCLC treatment, the emergence of drug resistance poses a significant challenge. Identifying drug-resistant cells and characterizing the resistance type is crucial for guiding clinical interventions in NSCLC. The homogeneity of drug-sensitive/resistant cancer cells presents a challenge in their identification as well as in distinguishing tumor slices. Organelles, pivotal for cellular function, exhibit notable variations in the microenvironment among diverse cell types. In this work, three organelle-targeting nanoparticles, composed of fluorescent probes and polydopamine particles, collectively formed PPTA-SA (an organelle-targeting sensor array) for imaging NSCLC cells and tumor slices. With a deep learning network, PPTA-SA could be used for identification of drug-resistant lung cells and tumors. The achieved identification accuracy for drug-resistant NSCLC cells and NSCLC tumor slices was more than 99%. Moreover, the multiorganelle targeting photothermal therapy demonstrated superior tumor ablation effects compared to conventional single-organelle targeting photothermal therapy. The combination of fluorescent probes and polydopamine not only served as a valuable tool for drug-sensitive/resistant NSCLC identification but also facilitated photothermal therapy with enhanced effects.

PMID:40762433 | DOI:10.1021/acs.analchem.5c02524

ACS Appl Mater Interfaces. 2025 Aug 5. doi: 10.1021/acsami.5c11459. Online ahead of print.

ABSTRACT

Fibrillar protein hydrogels are promising sustainable biomaterials for biomedical applications, but their practical use is often limited by insufficient mechanical strength and stability. To address these challenges, we transformed native proteins into amyloid fibrils (AFs) and incorporated a fibrillar polysaccharide, phytagel (PHY), to engineer interpenetrating polymer network (IPN) hydrogels. Notably, we report for the first time the formation of an amyloid-based hydrogel from apoferritin (APO), with PHY reinforcing the network's mechanical integrity. In situ self-assembly of APO within the PHY matrix yields fully natural, biopolymer-based IPNs. Rheological analyses confirm synergistic interactions between AF and PHY fibers, with the composite hydrogels exhibiting significantly enhanced viscoelastic moduli compared with individual components. The AF-PHY hydrogels also demonstrate excellent self-healing behavior, rapidly restoring their storage modulus after high-strain deformation. A major advancement of this study is the application of deep learning (DL)-based image analysis, using convolutional neural networks, to automate the identification, segmentation, and quantification of fibrillar components in high-resolution scanning electron microscopy images. This AI-driven method enables precise differentiation between AF and PHY fibers and reveals the three-dimensional microarchitecture of the IPN, overcoming key limitations of traditional image analysis. Complementary real-time confocal laser scanning microscopy, with selective fluorescent labeling of protein and polysaccharide components, further validates the IPN structure of the hybrid hydrogels. Our results demonstrate that DL significantly enhances structural characterization and provides insights into gelation processes. This approach sets a new guide for the analysis of complex soft materials and underlines the potential of AF-PHY hydrogels as mechanically robust, self-healing, and fully sustainable biomaterials for biomedical engineering applications.

PMID:40762431 | DOI:10.1021/acsami.5c11459

Proteins. 2025 Aug 5. doi: 10.1002/prot.70030. Online ahead of print.

ABSTRACT

Biomolecular structure prediction has reached an unprecedented level of accuracy, partly attributed to the use of advanced deep learning algorithms. We participated in the CASP16 experiments across the categories of protein domains, protein multimers, and RNA monomers, achieving official rankings of first, second, and fourth (top for server groups), respectively. We hypothesized that by leveraging state-of-the-art structure predictors such as AlphaFold2, AlphaFold3, trRosettaX2, and trRosettaRNA2, accurate structure predictions could be achieved through careful optimization of input information. For protein structure prediction, we enhanced the input sequences by removing intrinsically disordered regions, a simple yet effective approach that yielded accurate models for protein domains. However, fewer than 25% of the protein multimers were predicted with high quality. In RNA structure prediction, optimizing the secondary structure input for trRosettaRNA2 resulted in more accurate predictions than AlphaFold3. In summary, our prediction results in CASP16 indicate that protein domain structure prediction has achieved high accuracy. However, predicting protein multimers and RNA structures remains challenging, and we anticipate new advancements in these areas in the coming years.

PMID:40762404 | DOI:10.1002/prot.70030

Magn Reson Med. 2025 Aug 5. doi: 10.1002/mrm.70013. Online ahead of print.

ABSTRACT

PURPOSE: Systematic study of the effects of averaging and other relevant training strategies in deep learning (DL)-based denoising is required to optimize such processing pipelines for improving the quality of arterial spin labeling (ASL) images.

METHODS: Different averaging strategies, including windowed and interleaved averaging methods, and different levels of averaging before and after convolutional neural network-based and transformer-based denoising were studied. The experiments were performed on 152 single-delay ASL scans from 152 subjects, including pulsed and pseudo-continuous ASL acquisitions. Four-fold cross-validation was implemented in all experiments. The effect of including calibration scans (M0) was studied and compared across images of different levels of signal-to-noise ratio (SNR). The generalizability of DL denoising was examined in experiments using low-SNR ground truth in training. The results were assessed using image-quality metrics including structural similarity, peak SNR, and normalized mean absolute error.

RESULTS: Including M0 was almost always beneficial, with a dependence on the SNR of the input ASL images. Windowed averaging outperformed interleaved averaging, supporting the practice of reducing scan time. Averaging of ASL images before DL denoising was more advantageous than averaging after. Matching the SNR levels of the images in training and inferencing was important for optimal performance. These findings were consistent across convolutional neural network-based and transformer-based models. The generalizability of DL-based denoising was confirmed, and its capability to reduce artifacts was observed.

CONCLUSION: This study supports the use of DL-based denoising in improving the image quality of ASL and reducing scan time and provides insights to help optimize DL-denoising pipelines.

PMID:40762194 | DOI:10.1002/mrm.70013

RSC Adv. 2025 Aug 4;15(34):27795-27815. doi: 10.1039/d5ra03747f. eCollection 2025 Aug 1.

ABSTRACT

The integration of artificial intelligence (AI) in targeted anticancer drug delivery represents a significant advancement in oncology, offering innovative solutions to enhance the precision and effectiveness of cancer treatments. This review explores the various AI methodologies that are transforming the landscape of targeted drug delivery systems. By leveraging machine learning algorithms, researchers can analyze extensive datasets, including genomic, proteomic, and clinical data, to identify patient-specific factors that influence therapeutic responses. Supervised learning techniques, such as support vector machines and random forests, enable the classification of cancer types and the prediction of treatment outcomes based on historical data. Deep learning approaches, particularly convolutional neural networks, facilitate improved tumor detection and characterization through advanced imaging analysis. Moreover, reinforcement learning optimizes treatment protocols by dynamically adjusting drug dosages and administration schedules based on real-time patient responses. The convergence of AI and targeted anticancer drug delivery holds the promise of advancing cancer therapy by providing tailored treatment strategies that enhance efficacy while minimizing side effects. By improving the understanding of tumor biology and patient variability, AI-driven methods can facilitate the transition from traditional treatment paradigms to more personalized and effective cancer care. This review discusses the challenges and limitations of implementing AI in targeted anticancer drug delivery, including data quality, interpretability of AI models, and the need for robust validation in clinical settings.

PMID:40761897 | PMC:PMC12320933 | DOI:10.1039/d5ra03747f

Ther Adv Musculoskelet Dis. 2025 Aug 1;17:1759720X251357188. doi: 10.1177/1759720X251357188. eCollection 2025.

ABSTRACT

While severe vasculopathic manifestations of systemic sclerosis (SSc) are well-recognized, characterization of subclinical progressive vasculopathy contributing to cardiac involvement remains an unmet clinical need. This review highlights the evolving understanding of SSc heart involvement (SHI), including current standard clinical cardiac evaluation methods, prevalence of various cardiac manifestations of SHI, and advances at the forefront of precision medicine. Informed by this growing body of literature, we describe the development of a novel interdisciplinary cardio-rheumatology clinic at the Vanderbilt University Medical Center. Utilizing advances in imaging techniques and systemic retrieval and analysis of complex data sets, our dedicated cardio-rheumatology clinic offers opportunities for therapeutic advances and personalized medicine through mechanistic disease phenotyping in SSc. Nailfold capillaroscopy, thermography, and hand ultrasound with Doppler are acquired to characterize small vessel vasculopathy, while echocardiogram, ambulatory cardiac rhythm monitoring, cardiac magnetic resonance imaging, and cardiac positron emission tomography/computed tomography are utilized to characterize cardiac disease. By correlating vasculopathy imaging with cardiac manifestations, our cardio-rheumatology clinic aims to identify patients with SSc who would benefit from additional cardiac investigation even in the absence of cardiac symptomatology. This interdisciplinary collaboration may allow earlier detection of primary SHI, which is a common cause of death in SSc patients, resulting from both morpho-functional and electrical cardiac abnormalities. Our shared model of care and robust data acquisition facilitate clinical investigation by utilizing technological advances in data management. Using deep learning and pattern recognition, artificial intelligence (AI) offers opportunities to integrate data from imaging and monitoring techniques outlined in this report to provide quantifiable markers of disease progression and treatment efficacy. Given the potential for extensive AI data processing but the low prevalence of SSc, developing a multicenter cloud-based image sharing platform would accelerate clinical investigation in the field. Ultimately, we aim to tailor therapeutic decisions and risk mitigation strategies to improve SSc patient outcomes.

PMID:40761822 | PMC:PMC12319188 | DOI:10.1177/1759720X251357188

Front Bioinform. 2025 Jul 21;5:1603133. doi: 10.3389/fbinf.2025.1603133. eCollection 2025.

ABSTRACT

INTRODUCTION: Diabetes Mellitus (DM) constitutes a global epidemic and is one of the top ten leading causes of mortality (WHO, 2019), projected to rank seventh by 2030. The US National Diabetes Statistics Report (2021) states that 38.4 million Americans have diabetes. Dipeptidyl Peptidase-4 (DPP-4) is an FDA-approved target for the treatment of type 2 diabetes mellitus (T2DM). However, current DPP-4 inhibitors may cause adverse effects, including gastrointestinal issues, severe joint pain (FDA safety warning), nasopharyngitis, hypersensitivity, and nausea. Moreover, the development of novel drugs and the in vivo assessment of DPP-4 inhibition are both costly and often impractical. These challenges highlight the urgent need for efficient in-silico approaches to facilitate the discovery and optimization of safer and more effective DPP-4 inhibitors.

METHODOLOGY: Quantitative Structure-Activity Relationship (QSAR) modeling is a widely used computational approach for evaluating the properties of chemical substances. In this study, we employed a Neuro-symbolic (NeSy) approach, specifically the Logic Tensor Network (LTN), to develop a DPP-4 QSAR model capable of identifying potential small-molecule inhibitors and predicting bioactivity classification. For comparison, we also implemented baseline models using Deep Neural Networks (DNNs) and Transformers. A total of 6,563 bioactivity records (SMILES-based compounds with IC50 values) were collected from ChEMBL, PubChem, BindingDB, and GTP. Feature sets used for model training included descriptors (CDK Extended-PaDEL), fingerprints (Morgan), chemical language model embeddings (ChemBERTa-2), LLaMa 3.2 embedding features, and physicochemical properties.

RESULTS: Among all tested configurations, the Neuro-symbolic QSAR model (NeSyDPP-4) performed best using a combination of CDK extended and Morgan fingerprints. The model achieved an accuracy of 0.9725, an F1-score of 0.9723, an ROC AUC of 0.9719, and a Matthews correlation coefficient (MCC) of 0.9446. These results outperformed the baseline DNN and Transformer models, as well as existing state-of-the-art (SOTA) methods. To further validate the robustness of the model, we conducted an external evaluation using the Drug Target Common (DTC) dataset, where NeSyDPP-4 also demonstrated strong performance, with an accuracy of 0.9579, an AUC-ROC of 0.9565, a Matthews Correlation Coefficient (MCC) of 0.9171, and an F1-score of 0.9577.

DISCUSSION: These findings suggest that the NeSyDPP-4 model not only delivered high predictive performance but also demonstrated generalizability to external datasets. This approach presents a cost-effective and reliable alternative to traditional vivo screening, offering valuable support for the identification and classification of biologically active DPP-4 inhibitors in the treatment of type 2 diabetes mellitus (T2DM).

PMID:40761758 | PMC:PMC12319772 | DOI:10.3389/fbinf.2025.1603133

Front Plant Sci. 2025 Jul 21;16:1645490. doi: 10.3389/fpls.2025.1645490. eCollection 2025.

ABSTRACT

INTRODUCTION: Microplastics (MPs), as emerging environmental contaminants, pose a significant threat to global food security. In order to rapidly screen and diagnosis rice seedling under MPs stress at an early stage, it is essential to develop efficient and non-destructive detection methods.

METHODS: In this study, rice seedlings exposed to different concentrations (0, 10, and 100 mg/L) of polyethylene terephthalate (PET), polystyrene (PS), and polyvinyl chloride (PVC) MPs stress were constructed. Two complementary spectroscopic techniques, visible/near-infrared hyperspectral imaging (VNIR-HSI) and synchrotron radiation-based Fourier Transform Infrared spectroscopy (SR-FTIR), were employed to capture the biochemical changes of leaf organic molecules.

RESULTS: The spectral information of rice seedlings under MPs stress was obtained by using VNIR-HSI, and the low-dimensional clustering distribution analysis of the original spectra was conducted. An improved SE-LSTM full-spectral detection model was proposed, and the detection accuracy rate was greater than 93.88%. Characteristic wavelengths were extracted to build a simplified detection model, and the SHapley Additive exPlanations (SHAP) framework was applied to interpret the model by identifying the bands associated with chlorophyll, carotenoids, water content, and cellulose. Meanwhile, SR-FTIR spectroscopy was used to investigate compositional changes in both leaf lamina and veins, and two-dimensional correlation spectroscopy (2DCOS) was employed to reveal the sequential interactions among molecular components.

DISCUSSION: In conclusion, the combination of spectral technology and deep learning to capture the physiological and biochemical reactions of leaves could provide a rapid and interpretable method for detecting rice seedlings under MPs stress. This method could provide a solution for the early detection of external stress on other crops.

PMID:40761567 | PMC:PMC12318996 | DOI:10.3389/fpls.2025.1645490

Front Plant Sci. 2025 Jul 21;16:1640806. doi: 10.3389/fpls.2025.1640806. eCollection 2025.

ABSTRACT

INTRODUCTION: Accurate wheat yield estimation is crucial for efficient crop management. This study introduces the Spatio-Temporal Fusion Mixture of Experts (STF-MoE) model, an innovative deep learning framework built upon an LSTM-Transformer architecture.

METHODS: The STF-MoE model incorporates a heterogeneous Mixture of Experts (MoE) mechanism with an adaptive gating network. This design dynamically processes fused multi-source remote sensing features (e.g., near-infrared vegetation reflectance, NIRv; fraction of photosynthetically active radiation absorption, Fpar) and environmental variables (e.g., relative humidity, digital elevation model) across multiple expert networks. The model was applied to estimate wheat yield in six major Chinese provinces.

RESULTS: The STF-MoE model demonstrated exceptional accuracy in the most recent estimation year (R² = 0.827, RMSE = 547.7 kg/ha) and exhibited robust performance across historical years and extreme climatic events, outperforming baseline models. Relative humidity and digital elevation model were identified as the most critical yield-influencing factors. Furthermore, the model accurately estimated yield 1-2 months before harvest by identifying key phenological stages (March to June).

DISCUSSION: STF-MoE effectively handles multi-source spatiotemporal complexity via its dynamic gating and expert specialization. While underestimation persists in extreme-yield regions, the model provides a scalable solution for pre-harvest yield estimation. Future work will optimize computational efficiency and integrate higher-resolution data.

PMID:40761564 | PMC:PMC12318938 | DOI:10.3389/fpls.2025.1640806

Front Plant Sci. 2025 Jul 21;16:1583526. doi: 10.3389/fpls.2025.1583526. eCollection 2025.

ABSTRACT

Pod numbers are important for assessing soybean yield. How to simplify the traditional manual process and determine the pod number phenotype of soybean maturity more quickly and accurately is an urgent challenge for breeders. With the development of smart agriculture, numerous scientists have explored the phenotypic information related to soybean pod number and proposed corresponding methods. However, these methods mainly focus on the total number of pods, ignoring the differences between different pod types and do not consider the time-consuming and labor-intensive problem of picking pods from the whole plant. In this study, a deep learning approach was used to directly detect the number of different types of pods on non-disassembled plants at the maturity stage of soybean. Subsequently, the number of pods wascorrected by means of a metric learning method, thereby improving the accuracy of counting different types of pods. After 200 epochs, the recognition results of various object detection algorithms were compared to obtain the optimal model. Among the algorithms, YOLOX exhibited the highest mean average precision (mAP) of 83.43% in accurately determining the counts of diverse pod categories within soybean plants. By improving the Siamese Network in metric learning, the optimal Siamese Network model was obtained. SE-ResNet50 was used as the feature extraction network, and its accuracy on the test set reached 93.7%. Through the Siamese Network model, the results of object detection were further corrected and counted. The correlation coefficients between the number of one-seed pods, the number of two-seed pods, the number of three-seed pods, the number of four-seed pods and the total number of pods extracted by the algorithm and the manual measurement results were 92.62%, 95.17%, 96.90%, 94.93%, 96.64%,respectively. Compared with the object detection algorithm, the recognition of soybean mature pods was greatly improved, evolving into a high-throughput and universally applicable method. The described results show that the proposed method is a robust measurement and counting algorithm, which can reduce labor intensity, improve efficiency and accelerate the process of soybean breeding.

PMID:40761559 | PMC:PMC12319039 | DOI:10.3389/fpls.2025.1583526

Data Brief. 2025 Jul 11;61:111867. doi: 10.1016/j.dib.2025.111867. eCollection 2025 Aug.

ABSTRACT

Beaches are recognized as major sinks of plastic litter and key sites where litter fragments into countless small pieces. Because those fine particles are almost impossible to remove from the natural environment, it is essential to monitor macroplastic litter on beaches before they degrade. To observe the distribution of this litter in detail, it is essential to have automated and objective image-processing methods that can be applied to images captured by remote sensing devices, such as web cameras and drones. To develop such an automated analysis method, a deep learning-based approach has recently become mainstream, and clarifying technical issues based on case studies is vital. The preparation of training data for those practices is critical but laborious. The BePLi Dataset v2 is updated from BePLi Dataset v1, comprises 3722 original images of beach plastic litter and 118,572 manually processed annotations. All original images were obtained from the natural coastal environment on the Northwest Japan coast, and annotation for plastic litter was provided at both the pixel and individual levels. The plastic litter objects are categorized into thirteen representative plastic object classes: "pet_bottle," "other_bottle," "plastic_bag," "box_shaped_case," "other_container," "rope," "other_string," "fishing_net," "buoy," "other_fishing_gear," "styrene_foam," "others" and "fragment." The BePLi Dataset v2 allows users to develop an instance segmentation and object detection method detecting macro beach plastic litter individually and at the pixel level. Depending on the user, this dataset can serve multiple purposes at different levels of technology development, from counting objects to estimating litter coverage, as it provides both bounding box- and pixel-based annotations.

PMID:40761540 | PMC:PMC12320089 | DOI:10.1016/j.dib.2025.111867

Bioimpacts. 2025 Jun 28;15:31226. doi: 10.34172/bi.31226. eCollection 2025.

ABSTRACT

INTRODUCTION: Identifying gene-disease associations is crucial for advancing medical research and improving clinical outcomes. Nevertheless, the rapid expansion of biomedical literature poses significant obstacles to extracting meaningful relationships from extensive text collections.

METHODS: This study uses deep learning techniques to automate this process, using publicly available datasets (EU-ADR, GAD, and SNPPhenA) to classify these associations accurately. Each dataset underwent rigorous pre-processing, including entity identification and preparation, word embedding using pre-trained Word2Vec and fastText models, and position embedding to capture semantic and contextual relationships within the text. In this research, three deep learning-based hybrid models have been implemented and contrasted, including CNN-LSTM, CNN-GRU, and CNN-GRU-LSTM. Each model has been equipped with attentional mechanisms to enhance its performance.

RESULTS: Our findings reveal that the CNN-GRU model achieved the highest accuracy of 91.23% on the SNPPhenA dataset, while the CNN-GRU-LSTM model attained an accuracy of 90.14% on the EU-ADR dataset. Meanwhile, the CNN-LSTM model demonstrated superior performance on the GAD dataset, achieving an accuracy of 84.90%. Compared to previous state-of-the-art methods, such as BioBERT-based models, our hybrid approach demonstrates superior classification performance by effectively capturing local and sequential features without relying on heavy pre-training.

CONCLUSION: The developed models and their evaluation data are available at https://github.com/NoorFadhil/Deep-GDAE.

PMID:40761527 | PMC:PMC12319213 | DOI:10.34172/bi.31226

Cogn Neurodyn. 2025 Dec;19(1):123. doi: 10.1007/s11571-025-10304-3. Epub 2025 Aug 2.

ABSTRACT

Emotion recognition is a difficult problem mainly because emotions are presented in different modalities including; speech, face, and text. In light of this, in this paper, we introduce a novel framework known as Audio, Visual, and Text Emotions Fusion Network that will enhance the approaches to analyzing emotions that can incorporate these dissimilar types of inputs efficiently for the enhancement of the existing approaches to analyzing emotions. Using specialized techniques, each modality in this framework shows Graph Attention Network-based Transformer Network by employing Graph Attention Networks to detect dependencies in facial regions; Hybrid Wav2Vec 2.0 and Convolutional Neural Network combines Wav2Vec 2.0, and Convolutional Neural Network to extract informative temporal and frequency domain audio features. Contextual and sequential text semantics are captured by Bidirectional Encoder Representations from Transformers with Bidirectional Gated Recurrent Unit. They are fused based on a novel attention-based mechanism that distributes weights depending on the emotional context and improves cross-modal interactions. Moreover, the Audio, Visual, and Text Emotions Fusion Network system effectively identifies emotions, and the result section that contains overall accuracy at 98.7%, precision at 98.2%, recall, at 97.2%, and F1-score of 97.49% makes the proposed approach strong and efficient for real-time emotion recognition strategies.

PMID:40761311 | PMC:PMC12317966 | DOI:10.1007/s11571-025-10304-3