Eur J Nucl Med Mol Imaging. 2024 Dec 23. doi: 10.1007/s00259-024-07024-x. Online ahead of print.

ABSTRACT

PURPOSE: Extranodal natural killer/T-cell lymphoma (ENKTCL) is an hematologic malignancy with prognostic heterogeneity. We aimed to develop and validate DeepENKTCL, an interpretable deep learning prediction system for prognosis risk stratification in ENKTCL.

METHODS: A total of 562 patients from four centers were divided into the training cohort, validation cohort and test cohort. DeepENKTCL combined a tumor segmentation model, a PET/CT fusion model, and prognostic prediction models. RadScore and TopoScore were constructed using radiomics and topological features derived from fused images, with SHapley Additive exPlanations (SHAP) analysis enhancing interpretability. The final prognostic models, termed FusionScore, were developed for predicting progression-free survival (PFS) and overall survival (OS). Performance was assessed using area under the receiver operator characteristic curve (AUC), time-dependent C-index, clinical decision curves (DCA), and Kaplan-Meier (KM) curves.

RESULTS: The tumor segmentation model accurately delineated the tumor lesions. RadScore (AUC: 0.908 for PFS, 0.922 for OS in validation; 0.822 for PFS, 0.867 for OS in test) and TopoScore (AUC: 0.756 for PFS, 0.805 for OS in validation; 0.689 for PFS, 0.769 for OS in test) both exhibited potential prognostic capability. The time-dependent C-index (0.897 for PFS, 0.928 for OS in validation; 0.894 for PFS, 0.868 for OS in test) and DCA indicated that FusionScore offers significant prognostic performance and superior net clinical benefits compared to existing models. KM survival analysis showed that higher FusionScores correlated with poorer PFS and OS across all cohorts.

CONCLUSION: DeepENKTCL provided a robust and interpretable framework for ENKTCL prognosis, with the potential to improve patient outcomes and guide personalized treatment.

PMID:39714634 | DOI:10.1007/s00259-024-07024-x

Interdiscip Sci. 2024 Dec 23. doi: 10.1007/s12539-024-00673-4. Online ahead of print.

ABSTRACT

Cell-Penetrating Peptides (CPPs) are a crucial carrier for drug delivery. Since the process of synthesizing new CPPs in the laboratory is both time- and resource-consuming, computational methods to predict potential CPPs can be used to find CPPs to enhance the development of CPPs in therapy. In this study, EnDM-CPP is proposed, which combines machine learning algorithms (SVM and CatBoost) with convolutional neural networks (CNN and TextCNN). For dataset construction, three previous CPP benchmark datasets, including CPPsite 2.0, MLCPP 2.0, and CPP924, are merged to improve the diversity and reduce homology. For feature generation, two language model-based features obtained from the Transformer architecture, including ProtT5 and ESM-2, are employed in CNN and TextCNN. Additionally, sequence features, such as CPRS, Hybrid PseAAC, KSC, etc., are input to SVM and CatBoost. Based on the result of each predictor, Logistic Regression (LR) is built to predict the final decision. The experiment results indicate that ProtT5 and ESM-2 fusion features significantly contribute to predicting CPP and that combining employed features and models demonstrates better association. On an independent test dataset comparison, EnDM-CPP achieved an accuracy of 0.9495 and a Matthews correlation coefficient of 0.9008 with an improvement of 2.23%-9.48% and 4.32%-19.02%, respectively, compared with other state-of-the-art methods. Code and data are available at https://github.com/tudou1231/EnDM-CPP.git .

PMID:39714579 | DOI:10.1007/s12539-024-00673-4

Mol Divers. 2024 Dec 23. doi: 10.1007/s11030-024-11044-y. Online ahead of print.

ABSTRACT

Protein-ligand interactions are the molecular basis of many important cellular activities, such as gene regulation, cell metabolism, and signal transduction. Protein-ligand binding affinity is a crucial metric of the strength of the interaction between the two, and accurate prediction of its binding affinity is essential for discovering drugs' new uses. So far, although many predictive models based on machine learning and deep learning have been reported, most of the models mainly focus on one-dimensional sequence and two-dimensional structural characteristics of proteins and ligands, but fail to deeply explore the detailed interaction information between proteins and ligand atoms in the binding pocket region of three-dimensional space. In this study, we introduced a novel 4D tensor feature to capture key interactions within the binding pocket and developed a three-dimensional convolutional neural network (CNN) model based on this feature. Using ten-fold cross-validation, we identified the optimal parameter combination and pocket size. Additionally, we employed feature engineering to extract features across multiple dimensions, including one-dimensional sequences, two-dimensional structures of the ligand and protein, and three-dimensional interaction features between them. We proposed an efficient protein-ligand binding affinity prediction model MCDTA (multi-dimensional convolutional drug-target affinity), built on a multi-dimensional convolutional neural network framework. Feature ablation experiments revealed that the 4D tensor feature had the most significant impact on model performance. MCDTA performed exceptionally well on the PDBbind v.2020 dataset, achieving an RMSE of 1.231 and a PCC of 0.823. In comparative experiments, it outperformed five other mainstream binding affinity prediction models, with an RMSE of 1.349 and a PCC of 0.795. Moreover, MCDTA demonstrated strong generalization ability and practical screening performance across multiple benchmark datasets, highlighting its reliability and accuracy in predicting protein-ligand binding affinity. The code for MCDTA is available at https://github.com/dfhuang-AI/MCDTA .

PMID:39714563 | DOI:10.1007/s11030-024-11044-y

Med Phys. 2024 Dec 23. doi: 10.1002/mp.17581. Online ahead of print.

ABSTRACT

BACKGROUND: Measurement noise often leads to inaccurate shear wave phase velocity estimation in ultrasound shear wave elastography. Filtering techniques are commonly used for denoising the shear wavefields. However, these filters are often not sufficient, especially in fatty tissues where the signal-to-noise ratio (SNR) can be very low.

PURPOSE: The purpose of this study is to develop a deep learning approach for denoising shear wavefields in ultrasound shear wave elastography. This may lead to improved reconstruction of shear wave phase velocity image maps.

METHODS: The study addresses noise by transforming particle velocity data into a time-frequency representation. A neural network with encoder and decoder convolutional blocks effectively decomposes the input and extracts the signal of interest, improving the SNR in high-noise scenarios. The network is trained on simulated phantoms with elasticity values ranging from 3 to 60 kPa. A total of 1 85 570 samples with 80%-20 % $\%$ split were used for training and validation. The approach is tested on experimental phantom and ex-vivo goat liver tissue data. Performance was compared with the traditional filtering methods such as bandpass, median, and wavelet filtering. Kruskal-Wallis one-way analysis of variance was performed to check statistical significance. Multiple comparisons were performed using the Mann-Whitney U test and Holm-Bonferroni adjustment of p - values $p-{\rm values}$ .

RESULTS: The results are evaluated using SNR and the percentage of pixels that can be reconstructed in the phase velocity maps. The SNR levels in experimental data improved from -2 to 9.9 dB levels to 15.6 to 30.3 dB levels. Kruskal-Wallis one-way analysis showed statistical significance ( p < 0.05 $p<0.05$ ). Multiple comparisons with p-value corrections also showed statistically significant improvement when compared to the bandpass and wavelet filtering scheme ( p < 0.05 $p<0.05$ ). Smoother phase velocity maps were reconstructed after denoising. The coefficient of variation is less than 5 % $5\%$ in CIRS phantom and less than 18 % $18\%$ in ex-vivo goat liver tissue.

CONCLUSIONS: The proposed approach demonstrates improvement in shear wave phase velocity image map reconstruction and holds promise that deep learning methods can be effectively utilized to extract true shear wave signal from measured noisy data.

PMID:39714072 | DOI:10.1002/mp.17581

Med Phys. 2024 Dec 23. doi: 10.1002/mp.17600. Online ahead of print.

ABSTRACT

BACKGROUND: Gadolinium-based contrast agents (GBCAs) are commonly used in MRI scans of patients with gliomas to enhance brain tumor characterization using T1-weighted (T1W) MRI. However, there is growing concern about GBCA toxicity. This study develops a deep-learning framework to generate T1-postcontrast (T1C) from pre-contrast multiparametric MRI.

PURPOSE: We propose the tumor-aware vision transformer (TA-ViT) model that predicts high-quality T1C images. The predicted tumor region is significantly improved (p < 0.001) by conditioning the transformer layers from predicted segmentation maps through the adaptive layer norm zero mechanism. The predicted segmentation maps were generated with the multi-parametric residual (MPR) ViT model and transformed into a latent space to produce compressed, feature-rich representations. The TA-ViT model was applied to T1w and T2-FLAIR to predict T1C MRI images of 501 glioma cases from an open-source dataset. Selected patients were split into training (N = 400), validation (N = 50), and test (N = 51) sets. Model performance was evaluated with the peak-signal-to-noise ratio (PSNR), normalized cross-correlation (NCC), and normalized mean squared error (NMSE).

RESULTS: Both qualitative and quantitative results demonstrate that the TA-ViT model performs superior against the benchmark MPR-ViT model. Our method produces synthetic T1C MRI with high soft tissue contrast and more accurately synthesizes both the tumor and whole brain volumes. The synthesized T1C images achieved remarkable improvements in both tumor and healthy tissue regions compared to the MPR-ViT model. For healthy tissue and tumor regions, the results were as follows: NMSE: 8.53 ± 4.61E-4; PSNR: 31.2 ± 2.2; NCC: 0.908 ± 0.041 and NMSE: 1.22 ± 1.27E-4, PSNR: 41.3 ± 4.7, and NCC: 0.879 ± 0.042, respectively.

CONCLUSION: The proposed method generates synthetic T1C images that closely resemble real T1C images. Future development and application of this approach may enable contrast-agent-free MRI for brain tumor patients, eliminating the risk of GBCA toxicity and simplifying the MRI scan protocol.

PMID:39714049 | DOI:10.1002/mp.17600

ArXiv [Preprint]. 2024 Dec 10:arXiv:2412.07136v1.

ABSTRACT

PURPOSE: A reliable cancer prognosis model for clear cell renal cell carcinoma (ccRCC) can enhance personalized treatment. We developed a multi-modal ensemble model (MMEM) that integrates pretreatment clinical data, multi-omics data, and histopathology whole slide image (WSI) data to predict overall survival (OS) and disease-free survival (DFS) for ccRCC patients.

METHODS: We analyzed 226 patients from The Cancer Genome Atlas Kidney Renal Clear Cell Carcinoma (TCGA-KIRC) dataset, which includes OS, DFS follow-up data, and five data modalities: clinical data, WSIs, and three multi-omics datasets (mRNA, miRNA, and DNA methylation). Separate survival models were built for OS and DFS. Cox-proportional hazards (CPH) model with forward feature selection is used for clinical and multi-omics data. Features from WSIs were extracted using ResNet and three general-purpose foundation models. A deep learning-based CPH model predicted survival using encoded WSI features. Risk scores from all models were combined based on training performance.

RESULTS: Performance was assessed using concordance index (C-index) and AUROC. The clinical feature-based CPH model received the highest weight for both OS and DFS tasks. Among WSI-based models, the general-purpose foundation model (UNI) achieved the best performance. The final MMEM model surpassed single-modality models, achieving C-indices of 0.820 (OS) and 0.833 (DFS), and AUROC values of 0.831 (3-year patient death) and 0.862 (cancer recurrence). Using predicted risk medians to stratify high- and low-risk groups, log-rank tests showed improved performance in both OS and DFS compared to single-modality models.

CONCLUSION: MMEM is the first multi-modal model for ccRCC patients, integrating five data modalities. It outperformed single-modality models in prognostic ability and has the potential to assist in ccRCC patient management if independently validated.

PMID:39713797 | PMC:PMC11661283

J Med Imaging (Bellingham). 2024 Nov;11(6):064502. doi: 10.1117/1.JMI.11.6.064502. Epub 2024 Dec 20.

ABSTRACT

PURPOSE: Various brain atlases are available to parcellate and analyze brain connections. Most traditional machine learning and deep learning studies analyzing Attention Deficit Hyperactivity Disorder (ADHD) have used either one or two brain atlases for their analysis. However, there is a lack of comprehensive research evaluating the impact of different brain atlases and associated factors such as connectivity measures and dimension reduction techniques on ADHD diagnosis.

APPROACH: This paper proposes an efficient and robust multimodality model that investigates various brain atlases utilizing different parcellation strategies and scales. Thirty combinations of six brain atlases and five distinct machine learning classifiers with optimized hyperparameters are implemented to identify the most promising brain atlas for ADHD diagnosis. These outcomes are validated using the statistical Friedman test. To enhance comprehensiveness, the impact of three different connectivity measures, each representing unique facets of brain connectivity, is also analyzed. Considering the extensive complexity of brain interconnections, the effect of various dimension reduction techniques on classification performance and execution time is also analyzed. The final model is integrated with phenotypic data to create an efficient multimodal ADHD classification model.

RESULTS: Experimental results on the ADHD-200 dataset demonstrate a significant variation in classification performance introduced by each factor. The proposed model outperforms many state-of-the-art ADHD approaches and achieves an accuracy of 77.59%, an area under the curve (AUC) score of 77.25% and an F 1 -score of 75.43%.

CONCLUSIONS: The proposed model offers clear guidance for researchers, helping to standardize atlas selection and associated factors and improve the consistency and accuracy of ADHD studies for more reliable clinical applications.

PMID:39713730 | PMC:PMC11661636 | DOI:10.1117/1.JMI.11.6.064502

Int J Cardiol Congenit Heart Dis. 2022 May 21;9:100396. doi: 10.1016/j.ijcchd.2022.100396. eCollection 2022 Sep.

ABSTRACT

Sudden cardiac death (SCD) is one of the most frequent causes of death in adult patients with congenital heart disease (ACHD). Despite the rare frequency of its occurrence, the incident appears often when unexpected, and many affected patients had not been identified priorly. Data on predictors for SCD are limited since the total number of ACHD is low. As the cohort is heterogeneous, it is difficult to define uniform risk factors that apply to all ACHD. Complexity of the congenital heart disease appears to play a role, but other factors may also be relevant and have not been sufficiently identified yet. In current guidelines, recommendations are primarily based on data of patients without congenital heart disease. With the ATROPOS registry, we are aiming to identify reliable risk factors for SCD. The registry enables physicians globally to include patients with congenital heart disease who died of or survived SCD. After acquisition, the data will be compared to an age and complexity of disease matched cohort to perform a case-control analysis. Subsequently, a further analysis will be performed using deep learning algorithms with artificial intelligence to amplify the gathered information and find reliable risk factors.

PMID:39713548 | PMC:PMC11658112 | DOI:10.1016/j.ijcchd.2022.100396

bioRxiv [Preprint]. 2024 Dec 10:2024.12.07.627102. doi: 10.1101/2024.12.07.627102.

ABSTRACT

G protein-coupled receptors (GPCRs) remain a focal point of research due to their critical roles in cell signaling and their prominence as drug targets. However, directly linking drug efficacy to receptor-mediated activation of specific intracellular transducers and the resulting physiological outcomes remains challenging. It is unclear whether the enhanced therapeutic window of certain drugs - defined as the dose range that provides effective therapy with minimal side effects - stems from their low intrinsic efficacy across all signaling pathways or ligand bias, wherein specific transducer subtypes are preferentially activated in a given cellular system compared to a reference ligand. Accurately predicting safer compounds, whether through low intrinsic efficacy or ligand bias, would greatly advance drug development. While AI models hold promise for such predictions, the development of deep learning models capable of reliably forecasting GPCR ligands with defined bioactivities remains challenging, largely due to the limited availability of high-quality data. To address this, we pre-trained a model on receptor sequences and ligand datasets across all class A GPCRs, and then refined it to predict low-efficacy compounds or biased agonists for individual class A GPCRs. This was achieved using transfer learning and a neural network incorporating natural language processing of target sequences and receptor mutation effects on signaling. These two fine-tuned models-one for low-efficacy agonists and one for biased agonists-are available on demand for each class A GPCR and enable virtual screening of large chemical libraries, thereby facilitating the discovery of compounds with potentially improved safety profiles.

PMID:39713468 | PMC:PMC11661127 | DOI:10.1101/2024.12.07.627102

Photonix. 2024;5(1):40. doi: 10.1186/s43074-024-00155-2. Epub 2024 Dec 18.

ABSTRACT

Holography is an essential technique of generating three-dimensional images. Recently, quantum holography with undetected photons (QHUP) has emerged as a groundbreaking method capable of capturing complex amplitude images. Despite its potential, the practical application of QHUP has been limited by susceptibility to phase disturbances, low interference visibility, and limited spatial resolution. Deep learning, recognized for its ability in processing complex data, holds significant promise in addressing these challenges. In this report, we present an ample advancement in QHUP achieved by harnessing the power of deep learning to extract images from single-shot holograms, resulting in vastly reduced noise and distortion, alongside a notable enhancement in spatial resolution. The proposed and demonstrated deep learning QHUP (DL-QHUP) methodology offers a transformative solution by delivering high-speed imaging, improved spatial resolution, and superior noise resilience, making it suitable for diverse applications across an array of research fields stretching from biomedical imaging to remote sensing. DL-QHUP signifies a crucial leap forward in the realm of holography, demonstrating its immense potential to revolutionize imaging capabilities and pave the way for advancements in various scientific disciplines. The integration of DL-QHUP promises to unlock new possibilities in imaging applications, transcending existing limitations and offering unparalleled performance in challenging environments.

SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s43074-024-00155-2.

PMID:39711807 | PMC:PMC11655614 | DOI:10.1186/s43074-024-00155-2

Digit Health. 2024 Dec 20;10:20552076241305168. doi: 10.1177/20552076241305168. eCollection 2024 Jan-Dec.

ABSTRACT

OBJECTIVE: While current multimodal approaches in the diagnosis and severity assessment of pneumonia demonstrate remarkable performance, they frequently overlook the issue of modality absence-a common challenge in clinical practice. Thus, we present the robust multimodal transformer (RMT) model, crafted to bridge this gap. The RMT model aims to enhance diagnosis and severity assessment accuracy in situations with incomplete data, thereby ensuring it meets the complex needs of real-world clinical settings.

METHOD: The RMT model leverages multimodal data, integrating X-ray images and clinical text data through a sophisticated AI-driven framework. It employs a Transformer-based architecture, enhanced by multi-task learning and mask attention mechanism. This approach aims to optimize the model's performance across different modalities, particularly under conditions of modality absence.

RESULTS: The RMT model demonstrates superior performance over traditional diagnostic methods and baseline models in accuracy, precision, sensitivity, and specificity. In tests involving various scenarios, including single-modal and multimodal tasks, the model shows remarkable robustness in handling incomplete data. Its effectiveness is further validated through extensive comparative analysis and ablation studies.

CONCLUSION: The RMT model represents a substantial advancement in pediatric pneumonia severity assessment. It successfully harnesses multimodal data and advanced AI techniques to improve assessment precision. While the RMT model sets a new precedent in AI applications in medical diagnostics, the development of a comprehensive pediatric pneumonia dataset marks a pivotal contribution, providing a robust foundation for future research.

PMID:39711742 | PMC:PMC11660274 | DOI:10.1177/20552076241305168

Digit Health. 2024 Dec 19;10:20552076241308987. doi: 10.1177/20552076241308987. eCollection 2024 Jan-Dec.

ABSTRACT

OBJECTIVE: The study aims to present an active learning approach that automatically extracts clinical concepts from unstructured data and classifies them into explicit categories such as Problem, Treatment, and Test while preserving high precision and recall and demonstrating the approach through experiments using i2b2 public datasets.

METHODS: Initially labeled data are acquired from a lexical-based approach in sufficient amounts to perform an active learning process. A contextual word embedding similarity approach is adopted using BERT base variant models such as ClinicalBERT, DistilBERT, and SCIBERT to automatically classify the unlabeled clinical concept into explicit categories. Additionally, deep learning and large language model (LLM) are trained on acquiring label data through active learning.

RESULTS: Using i2b2 datasets (426 clinical notes), the lexical-based method achieved precision, recall, and F1-scores of 76%, 70%, and 73%. SCIBERT excelled in active transfer learning, yielding precision of 70.84%, recall of 77.40%, F1-score of 73.97%, and accuracy of 69.30%, surpassing counterpart models. Among deep learning models, convolutional neural networks (CNNs) trained with embeddings (BERTBase, DistilBERT, SCIBERT, ClinicalBERT) achieved training accuracies of 92-95% and testing accuracies of 89-93%. These results were higher compared to other deep learning models. Additionally, we individually evaluated these LLMs; among them, ClinicalBERT achieved the highest performance, with a training accuracy of 98.4% and a testing accuracy of 96%, outperforming the others.

CONCLUSIONS: The proposed methodology enhances clinical concept extraction by integrating active learning and models like SCIBERT and CNN. It improves annotation efficiency while maintaining high accuracy, showcasing potential for clinical applications.

PMID:39711738 | PMC:PMC11660282 | DOI:10.1177/20552076241308987

medRxiv [Preprint]. 2024 Dec 10:2024.12.09.24318661. doi: 10.1101/2024.12.09.24318661.

ABSTRACT

BACKGROUND: Carotid atherosclerosis is a major etiology of stroke. Although intraplaque hemorrhage (IPH) is known to increase stroke risk and plaque burden, its long-term effects on plaque dynamics remain unclear.

OBJECTIVES: This study aimed to evaluate the long-term impact of IPH on carotid plaque burden progression using deep learning-based segmentation on multi-contrast vessel wall imaging (VWI).

METHODS: Twenty-eight asymptomatic subjects with carotid atherosclerosis underwent an average of 4.7 ± 0.6 VWI scans over 5.8 ± 1.1 years. Deep learning pipelines segmented the carotid vessel walls and IPH. Bilateral plaque progression was analyzed using generalized estimating equations, and linear mixed-effects models evaluated long-term associations between IPH occurrence, IPH volume (%HV), and plaque burden (%WV) progression.

RESULTS: Two subjects with ipsilateral IPH developed new ischemic infarcts during follow- up. IPH was detected in 23/50 of arteries. Of arteries without IPH at baseline, 11/39 developed new IPH that persisted, while 5/11 arteries with baseline IPH exhibited it throughout the study. Bilateral plaque growth was significantly correlated (r = 0.54, p < 0.001), but this symmetry was weakened with IPH presence. The progression rate for arteries without IPH was -0.001 %/year (p = 0.895). However, IPH presence or development at any point was associated with a 2.34% absolute increase in %WV (p < 0.001), and %HV was associated with 0.73% per 2-fold increase over the mean of %HV (p = 0.005).

CONCLUSIONS: IPH may persist asymptomatically for extended periods. While arteries without IPH demonstrated minimal progression under contemporary treatment, IPH significantly accelerated long-term plaque growth.

PMID:39711698 | PMC:PMC11661346 | DOI:10.1101/2024.12.09.24318661

Front Plant Sci. 2024 Dec 6;15:1475282. doi: 10.3389/fpls.2024.1475282. eCollection 2024.

ABSTRACT

INTRODUCTION: The assessment of the severity of fruit disease is crucial for the optimization of fruit production. By quantifying the percentage of leaf disease, an effective approach to determining the severity of the disease is available. However, the current prediction of disease degree by machine learning methods still faces challenges, including suboptimal accuracy and limited generalizability.

METHODS: In light of the growing application of large model technology across a range of fields, this study draws upon the DINOV2 visual large vision model backbone network to construct the DINOV2-Fruit Leaf Classification and Segmentation Model (DINOV2-FCS), a model designed for the classification and severity prediction of diverse fruit leaf diseases. DINOV2-FCS employs the DINOv2-B (distilled) backbone feature extraction network to enhance the extraction of features from fruit disease leaf images. In fruit leaf disease classification, for the problem that leaf spots of different diseases have great similarity, we have proposed Class-Patch Feature Fusion Module (C-PFFM), which integrates the local detailed feature information of the spots and the global feature information of the class markers. For the problem that the model ignores the fine spots in the segmentation process, we propose Explicit Feature Fusion Architecture (EFFA) and Alterable Kernel Atrous Spatial Pyramid Pooling (AKASPP), which improve the segmentation effect of the model.

RESULTS: To verify the accuracy and generalizability of the model, two sets of experiments were conducted. First, the labeled leaf disease dataset of five fruits was randomly divided. The trained model exhibited an accuracy of 99.67% in disease classification, an mIoU of 90.29%, and an accuracy of 95.68% in disease severity classification. In the generalizability experiment, four disease data sets were used for training and one for testing. The mIoU of the trained model reached 83.95%, and the accuracy of disease severity grading was 95.24%.

DISCUSSION: The results demonstrate that the model exhibits superior performance compared to other state-of-the-art models and that the model has strong generalization capabilities. This study provides a new method for leaf disease classification and leaf disease severity prediction for a variety of fruits. Code is available at https://github.com/BaiChunhui2001/DINOV2-FCS.

PMID:39711594 | PMC:PMC11658969 | DOI:10.3389/fpls.2024.1475282

Front Plant Sci. 2024 Dec 6;15:1423338. doi: 10.3389/fpls.2024.1423338. eCollection 2024.

ABSTRACT

It is crucial for robotic picking fruit to recognize fruit accurately in orchards, this paper reviews the applications and research results of target recognition in orchard fruit picking by using machine vision and emphasizes two methods of fruit recognition: the traditional digital image processing method and the target recognition method based on deep learning. Here, we outline the research achievements and progress of traditional digital image processing methods by the researchers aiming at different disturbance factors in orchards and summarize the shortcomings of traditional digital image processing methods. Then, we focus on the relevant contents of fruit target recognition methods based on deep learning, including the target recognition process, the preparation and classification of the dataset, and the research results of target recognition algorithms in classification, detection, segmentation, and compression acceleration of target recognition network models. Additionally, we summarize the shortcomings of current orchard fruit target recognition tasks from the perspectives of datasets, model applicability, universality of application scenarios, difficulty of recognition tasks, and stability of various algorithms, and look forward to the future development of orchard fruit target recognition.

PMID:39711588 | PMC:PMC11659763 | DOI:10.3389/fpls.2024.1423338

Front Plant Sci. 2024 Dec 6;15:1499911. doi: 10.3389/fpls.2024.1499911. eCollection 2024.

ABSTRACT

Accurate diagnosis of grape leaf diseases is critical in agricultural production, yet existing detection techniques face challenges in achieving model lightweighting while ensuring high accuracy. In this study, a real-time, end-to-end, lightweight grape leaf disease detection model, MHDI-DETR, based on an improved RT-DETR architecture, is presented to address these challenges. The original residual backbone network was improved using the MobileNetv4 network, significantly reducing the model's computational requirements and complexity. Additionally, a lightSFPN feature fusion structure is presented, combining the Hierarchical Scale Feature Pyramid Network with the Dilated Reparam Block structure design from the UniRepLKNet network. This structure is designed to overcome the challenges of capturing complex high-level and subtle low-level features, and it uses Efficient Local Attention to focus more efficiently on regions of interest, thereby enhancing the model's ability to detect complex targets while improving accuracy and inference speed. Finally, the integration of GIou and Focaler-IoU into Focaler-GIoU enhances detection accuracy and convergence speed for small targets by focusing more effectively on both simple and difficult samples. The findings from the experiments suggest that The MHDI-DETR model results in a 56% decrease in parameters and a 49% reduction in floating-point operations, respectively, compared with the RT-DETR model, in terms of accuracy, the model achieved precision rates of 96.9%, 92.6%, and 72.5% for accuracy, mAP50, and mAP50:95, respectively. Compared with the RT-DETR model, these represent improvements of 1.9%, 1.2%, and 1.2%. Overall, the MHDI-DETR model surpasses the RT-DETR and other mainstream detection models in both detection accuracy and degree of lightness, achieving dual optimization in efficiency and accuracy, and providing an efficient technical solution for automated agricultural disease management.

PMID:39711587 | PMC:PMC11659005 | DOI:10.3389/fpls.2024.1499911

Res Sq [Preprint]. 2024 Dec 13:rs.3.rs-5463126. doi: 10.21203/rs.3.rs-5463126/v1.

ABSTRACT

In the early stages of atrial fibrillation (AF), most cases are paroxysmal (pAF), making identification only possible with continuous and prolonged monitoring. With the advent of wearables, smartwatches equipped with photoplethysmographic (PPG) sensors are an ideal approach for continuous monitoring of pAF. There have been numerous studies demonstrating successful capture of pAF events, especially using deep learning. However, deep learning requires a large amount of data and independent testing on diverse datasets, to ensure the generalizability of the model, and most prior studies did not meet these requirements. Moreover, most prior studies using wearable-based PPG sensor data collection were limited either to controlled environments, to minimize motion artifacts, or to short duration data collection. Most importantly, frequent premature atrial and ventricular contractions (PAC/PVC) can confound most AF detection algorithms. This has not been well studied, largely due to limited datasets containing these rhythms. Note that the recent deep learning models show 97% AF detection accuracy, and the sensitivity of the current state-of-the-art technique for PAC/PVC detection is only 75% on minimally motion artifact corrupted PPG data. Our study aims to address the above limitations using a recently completed NIH-funded Pulsewatch clinical trial which collected smartwatch PPG data over two weeks from 106 subjects. For our approach, we used multi-modal data which included 1D PPG, accelerometer, and heart rate data. We used a computationally efficient 1D bi-directional Gated Recurrent Unit (1D-Bi-GRU) deep learning model to detect three classes: normal sinus rhythm, AF, and PAC/PVC. Our proposed 1D-Bi-GRU model's performance was compared with two other deep learning models that have reported some of the highest performance metrics, in prior work. For three-arrhythmia-classification, testing data for all deep learning models consisted of using independent data and subjects from the training data, and further evaluations were performed using two independent datasets that were not part of the training dataset. Our multimodal model achieved an unprecedented 83% sensitivity for PAC/PVC detection while maintaining a high accuracy of 97.31% for AF detection. Our model was computationally more efficient (14 times more efficient and 2.7 times faster) and outperformed the best state-of-the-art model by 20.81% for PAC/PVC sensitivity and 2.55% for AF accuracy. We also tested our models on two independent PPG datasets collected with a different smartwatch and a fingertip PPG sensor. Our three-arrhythmia-classification results show high macro-averaged area under the receiver operating characteristic curve values of 96.22%, and 94.17% for two independent datasets, demonstrating better generalizability of the proposed model.

PMID:39711547 | PMC:PMC11661413 | DOI:10.21203/rs.3.rs-5463126/v1

Res Sq [Preprint]. 2024 Dec 9:rs.3.rs-5334453. doi: 10.21203/rs.3.rs-5334453/v1.

ABSTRACT

Advances in magnetic resonance imaging (MRI) have revolutionized disease detection and treatment planning. However, as the volume and complexity of MRI data grow with increasing heterogeneity between institutions in imaging protocol, scanner technology, and data labeling, there is a need for a standardized methodology to efficiently identify, characterize, and label MRI sequences. Such a methodology is crucial for advancing research efforts that incorporate MRI data from diverse populations to develop robust machine learning models. This research utilizes convolutional neural networks (CNNs) to automatically classify sequence, orientation, and contrast, specifically tailored for abdominal MRI. Three distinct CNN models with similar backbone architectures were trained to classify single image slices into one of 12 sequences, 4 orientations, and 2 contrast classes. Results derived from this method demonstrate high levels of performance for the three specialized CNN models, with model accuracies for sequence, orientation, and contrast of 96.9%, 97.4%, and 97.3%, respectively.

PMID:39711527 | PMC:PMC11661311 | DOI:10.21203/rs.3.rs-5334453/v1

Anal Chem. 2024 Dec 23. doi: 10.1021/acs.analchem.4c04106. Online ahead of print.

ABSTRACT

The increasing focus of small extracellular vesicles (sEVs) in liquid biopsy has created a significant demand for streamlined improvements in sEV isolation methods, efficient collection of high-quality sEV data, and powerful rapid analysis of large data sets. Herein, we develop a high-throughput dual-use mass spectroscopic chip array (DUMSCA) for the rapid isolation and detection of plasma sEVs. The DUMSCA realizes more than a 50% increase in speed compared to traditional method and confirms proficiency in robust storage, reuse, high-efficiency desorption/ionization, and metabolite quantification. With the collected metabolic data matrix of sEVs, a deep learning model achieves high-performance diagnosis of Crohn's disease. Furthermore, discovered biomarkers by feature sparsification and tandem mass spectrometry experiments also exhibited remarkable performance in diagnosis. This work demonstrates the rapidity and validity of DUMSCA for disease diagnosis, enabling the diagnosis of diseases without the necessity for prior knowledge and providing a high-throughput technology for sEV-based liquid biopsy that will empower its vigorous development.

PMID:39711466 | DOI:10.1021/acs.analchem.4c04106

J Struct Biol. 2024 Dec 20:108163. doi: 10.1016/j.jsb.2024.108163. Online ahead of print.

ABSTRACT

Cryogenic electron tomography (cryo-ET) has rapidly advanced as a high-resolution imaging tool for visualizing subcellular structures in 3D with molecular detail. Direct image inspection remains challenging due to inherent low signal-to-noise ratios (SNR). We introduce CryoSamba, a self-supervised deep learning-based model designed for denoising cryo-ET images. CryoSamba enhances single consecutive 2D planes in tomograms by averaging motion-compensated nearby planes through deep learning interpolation, effectively mimicking increased exposure. This approach amplifies coherent signals and reduces high-frequency noise, substantially improving tomogram contrast and SNR. CryoSamba operates on 3D volumes without needing pre-recorded images, synthetic data, labels or annotations, noise models, or paired volumes. CryoSamba suppresses high-frequency information less aggressively than do existing cryo-ET denoising methods, while retaining real information, as shown both by visual inspection and by Fourier Shell Correlation (FSC) analysis of icosahedrally symmetric virus particles. Thus, CryoSamba enhances the analytical pipeline for direct 3D tomogram visual interpretation.

PMID:39710216 | DOI:10.1016/j.jsb.2024.108163