Med Phys. 2024 Oct 21. doi: 10.1002/mp.17466. Online ahead of print.

ABSTRACT

BACKGROUND: Deep learning-based computed tomography (CT) ventilation imaging (CTVI) is a promising technique for guiding functional lung avoidance radiotherapy (FLART). However, conventional approaches, which rely on anatomical CT data, may overlook important ventilation features due to the lack of motion data integration.

PURPOSE: This study aims to develop a novel dual-aware CTVI method that integrates both anatomical information from CT images and motional information from Jacobian maps to generate more accurate ventilation images for FLART.

METHODS: A dataset of 66 patients with four-dimensional CT (4DCT) images and reference ventilation images (RefVI) was utilized to develop the dual-path fusion network (DPFN) for synthesizing ventilation images (CTVIDual). The DPFN model was specifically designed to integrate motion data from 4DCT-generated Jacobian maps with anatomical data from average 4DCT images. The DPFN utilized two specialized feature extraction pathways, along with encoders and decoders, designed to handle both 3D average CT images and Jacobian map data. This dual-processing approach enabled the comprehensive extraction of lung ventilation-related features. The performance of DPFN was assessed by comparing CTVIDual to RefVI using various metrics, including Spearman's correlation coefficients (R), Dice similarity coefficients of high-functional region (DSCh), and low-functional region (DSCl). Additionally, CTVIDual was benchmarked against other CTVI methods, including a dual-phase CT-based deep learning method (CTVIDLCT), a radiomics-based method (CTVIFM), a super voxel-based method (CTVISVD), a Unet-based method (CTVIUnet), and two deformable registration-based methods (CTVIJac and CTVIHU).

RESULTS: In the test group, the mean R between CTVIDual and RefVI was 0.70, significantly outperforming CTVIDLCT (0.68), CTVIFM (0.58), CTVISVD (0.62), and CTVIUnet (0.66), with p < 0.05. Furthermore, the DSCh and DSCl values of CTVIDual were 0.64 and 0.80, respectively, outperforming CTVISVD (0.63; 0.73) and CTVIUnet (0.62; 0.77). The performance of CTVIDual was also significantly better than that of CTVIJac and CTVIHU.

CONCLUSIONS: A novel dual-aware CTVI model that integrates anatomical and motion information was developed to synthesize lung ventilation images. It was shown that the accuracy of lung ventilation estimation could be significantly enhanced by incorporating motional information, particularly in patients with tumor-induced blockages. This approach has the potential to improve the accuracy of CTVI, enabling more effective FLART.

PMID:39432032 | DOI:10.1002/mp.17466

Med Phys. 2024 Oct 21. doi: 10.1002/mp.17470. Online ahead of print.

ABSTRACT

BACKGROUND: Large foundation models, such as the Segment Anything Model (SAM), have shown remarkable performance in image segmentation tasks. However, the optimal approach to achieve true utility of these models for domain-specific applications, such as medical image segmentation, remains an open question. Recent studies have released a medical version of the foundation model MedSAM by training on vast medical data, who promised SOTA medical segmentation. Independent community inspection and dissection is needed.

PURPOSE: Foundation models are developed for general purposes. On the other hand, stable delivery of reliable performance is key to clinical utility. This study aims at elucidating the potential advantage and limitations of landing the foundation models in clinical use by assessing the performance of off-the-shelf medical foundation model MedSAM for the segmentation of anatomical structures in pelvic MR images. We also explore the simple remedies by evaluating the dependency on prompting scheme. Finally, we demonstrate the need and performance gain of further specialized fine-tuning.

METHODS: MedSAM and its lightweight version LiteMedSAM were evaluated out-of-the-box on a public MR dataset consisting of 589 pelvic images split 80:20 for training and testing. An nnU-Net model was trained from scratch to serve as a benchmark and to provide bounding box prompts for MedSAM. MedSAM was evaluated using different quality bounding boxes, those derived from ground truth labels, those derived from nnU-Net, and those derived from the former two but with 5-pixel isometric expansion. Lastly, LiteMedSAM was refined on the training set and reevaluated on this task.

RESULTS: Out-of-the-box MedSAM and LiteMedSAM both performed poorly across the structure set, especially for disjoint or non-convex structures. Varying prompt with different bounding box inputs had minimal effect. For example, the mean Dice score and mean Hausdorff distances (in mm) for obturator internus using MedSAM and LiteMedSAM were {0.251 ± 0.110, 0.101 ± 0.079} and {34.142 ± 5.196, 33.688 ± 5.306}, respectively. Fine-tuning of LiteMedSAM led to significant performance gain, improving Dice score and Hausdorff distance for the obturator internus to 0.864 ± 0.123 and 5.022 ± 10.684, on par with nnU-Net with no significant difference in evaluation of most structures. All segmentation structures benefited significantly from specialized refinement, at varying improvement margin.

CONCLUSION: While our study alludes to the potential of deep learning models like MedSAM and LiteMedSAM for medical segmentation, it highlights the need for specialized refinement and adjudication. Off-the-shelf use of such large foundation models is highly likely to be suboptimal, and specialized fine-tuning is often necessary to achieve clinical desired accuracy and stability.

PMID:39431952 | DOI:10.1002/mp.17470

ACS Sens. 2024 Oct 21. doi: 10.1021/acssensors.4c01097. Online ahead of print.

ABSTRACT

An omnidirectional stretchable strain sensor with high resolution is a critical component for motion detection and human-machine interaction. It is the current dominant solution to integrate several consistent units into the omnidirectional sensor based on a certain geometric structure. However, the excessive similarity in orientation characteristics among sensing units restricts orientation recognition due to their closely matched strain sensitivity. In this study, based on strain partition modulation (SPM), a sensitivity anisotropic amplification strategy is proposed for resistive strain sensors. The stress distribution of a sensitive conductive network is modulated by structural parameters of the customized periodic hole array introduced underneath the elastomer substrate. Meanwhile, the strain isolation structures are designed on both sides of the sensing unit for stress interference immune. The optimized sensors exhibit excellent sensitivity (19 for 0-80%; 109 for 80%-140%; 368 for 140%-200%), with nearly a 7-fold improvement in the 140%-200% interval compared to bare elastomer sensors. More importantly, a sensing array composed of multiple units with different hole configurations can highlight orientation characteristics with amplitude difference between channels reaching up to 29 times. For the 48-class strain-orientation decoupling task, the recognition rate of the sensitivity-differentiated layout sensor with the lightweight deep learning network is as high as 96.01%, superior to that of 85.7% for the sensitivity-consistent layout. Furthermore, the application of the sensor to the fitness field demonstrates an accurate recognition of the wrist flexion direction (98.4%) and spinal bending angle (83.4%). Looking forward, this methodology provides unique prospects for broader applications such as tactile sensors, soft robotics, and health monitoring technologies.

PMID:39431947 | DOI:10.1021/acssensors.4c01097

Rev Sci Instrum. 2024 Oct 1;95(10):103521. doi: 10.1063/5.0219526.

ABSTRACT

A deep-learning convolutional neural network (CNN) is used to infer, from x-ray images along multiple lines of sight, the low-mode shape of the hot-spot emission of deuterium-tritium (DT) laser-direct-drive cryogenic implosions on OMEGA. The motivation of this approach is to develop a physics-informed 3-D reconstruction technique that can be performed within minutes to facilitate the use of the results to inform changes to the initial target and laser conditions for the subsequent implosion. The CNN is trained on a 3D radiation-hydrodynamic simulation database to relate 2D x-ray images to 3D emissivity at stagnation. The CNN accounts for the lack of an absolute spatial reference and the different bands of photon energies in the x-ray images. While previous work [O. M. Mannion et al., Phys. Plasmas 28, 042701 (2021) and A. Lees et al., Phys. Rev. Lett. 127, 105001 (2021)] studied the effect of mode-1 asymmetries on implosion performance using nuclear diagnostics, this work focuses on the effect of mode 2 inferred from x-ray diagnostics on implosion performance. A current analysis of 19 DT cryogenic implosions indicates there is an upper limit of ∼20% reduction in the neutron yield caused by an ℓ = 2 amplitude for ℓ2/ℓ0 ≤ 0.32. These conclusions are supported by 2D simulations.

PMID:39431883 | DOI:10.1063/5.0219526

Pest Manag Sci. 2024 Oct 21. doi: 10.1002/ps.8456. Online ahead of print.

ABSTRACT

BACKGROUND: Cassava is a high-carbohydrate crop that is at risk of viral infections. The production rate and quality of cassava crops are affected by several diseases. However, the manual identification of diseases is challenging and requires considerable time because of the lack of field professionals and the limited availability of clear and distinct information. Consequently, the agricultural management system is seeking an efficient and lightweight method that can be deployable to edged devices for detecting diseases at an early stage. To address these issues and accurately categorize different diseases, a very effective and lightweight framework called CDDNet has been introduced. We used MobileNetV3Small framework as a backbone feature for extracting optimized, discriminating, and distinct features. These features are empirically validated at the early intermediate stage. Additionally, we modified the soft attention module to effectively prioritize the diseased regions and enhance significant cassava plant disease-related features for efficient cassava disease detection.

RESULTS: Our proposed method achieved accuracies of 98.95%, 97.03%, and 98.25% on Cassava Image Dataset, Cassava Plant Disease Merged (2019-2020) Dataset, and the newly created Cassava Plant Composite Dataset, respectively. Furthermore, the proposed technique outperforms previous state-of-the-art methods in terms of accuracy, parameter count, and frames per second values, ultimately making the proposed CDDNet the best one for real-time processing.

CONCLUSION: Our findings underscore the importance of a lightweight and efficient technique for cassava disease detection and classification in a real-time environment. Furthermore, we highlight the impact of modified soft attention on model performance. © 2024 The Author(s). Pest Management Science published by John Wiley & Sons Ltd on behalf of Society of Chemical Industry.

PMID:39431752 | DOI:10.1002/ps.8456

Brief Bioinform. 2024 Sep 23;25(6):bbae519. doi: 10.1093/bib/bbae519.

ABSTRACT

Protein degradation through the ubiquitin proteasome system at the spatial and temporal regulation is essential for many cellular processes. E3 ligases and degradation signals (degrons), the sequences they recognize in the target proteins, are key parts of the ubiquitin-mediated proteolysis, and their interactions determine the degradation specificity and maintain cellular homeostasis. To date, only a limited number of targeted degron instances have been identified, and their properties are not yet fully characterized. To tackle on this challenge, here we develop a novel deep-learning framework, namely MetaDegron, for predicting E3 ligase targeted degron by integrating the protein language model and comprehensive featurization strategies. Through extensive evaluations using benchmark datasets and comparison with existing method, such as Degpred, we demonstrate the superior performance of MetaDegron. Among functional features, MetaDegron allows batch prediction of targeted degrons of 21 E3 ligases, and provides functional annotations and visualization of multiple degron-related structural and physicochemical features. MetaDegron is freely available at http://modinfor.com/MetaDegron/. We anticipate that MetaDegron will serve as a useful tool for the clinical and translational community to elucidate the mechanisms of regulation of protein homeostasis, cancer research, and drug development.

PMID:39431517 | DOI:10.1093/bib/bbae519

Brief Bioinform. 2024 Sep 23;25(6):bbae517. doi: 10.1093/bib/bbae517.

ABSTRACT

Aptamers are single-stranded nucleic acid ligands, featuring high affinity and specificity to target molecules. Traditionally they are identified from large DNA/RNA libraries using $in vitro$ methods, like Systematic Evolution of Ligands by Exponential Enrichment (SELEX). However, these libraries capture only a small fraction of theoretical sequence space, and various aptamer candidates are constrained by actual sequencing capabilities from the experiment. Addressing this, we proposed AptaDiff, the first in silico aptamer design and optimization method based on the diffusion model. Our Aptadiff can generate aptamers beyond the constraints of high-throughput sequencing data, leveraging motif-dependent latent embeddings from variational autoencoder, and can optimize aptamers by affinity-guided aptamer generation according to Bayesian optimization. Comparative evaluations revealed AptaDiff's superiority over existing aptamer generation methods in terms of quality and fidelity across four high-throughput screening data targeting distinct proteins. Moreover, surface plasmon resonance experiments were conducted to validate the binding affinity of aptamers generated through Bayesian optimization for two target proteins. The results unveiled a significant boost of $87.9\%$ and $60.2\%$ in RU values, along with a 3.6-fold and 2.4-fold decrease in KD values for the respective target proteins. Notably, the optimized aptamers demonstrated superior binding affinity compared to top experimental candidates selected through SELEX, underscoring the promising outcomes of our AptaDiff in accelerating the discovery of superior aptamers.

PMID:39431516 | DOI:10.1093/bib/bbae517

Fundam Res. 2023 Jul 1;4(5):981-991. doi: 10.1016/j.fmre.2023.06.006. eCollection 2024 Sep.

ABSTRACT

Scene segmentation is widely used in autonomous driving for environmental perception. Semantic scene segmentation has gained considerable attention owing to its rich semantic information. It assigns labels to the pixels in an image, thereby enabling automatic image labeling. Current approaches are based mainly on convolutional neural networks (CNN), however, they rely on numerous labels. Therefore, the use of a small amount of labeled data to achieve semantic segmentation has become increasingly important. In this study, we developed a domain adaptation framework based on optimal transport (OT) and an attention mechanism to address this issue. Specifically, we first generated the output space via a CNN owing to its superior of feature representation. Second, we utilized OT to achieve a more robust alignment of the source and target domains in the output space, where the OT plan defined a well attention mechanism to improve the adaptation of the model. In particular, the OT reduced the number of network parameters and made the network more interpretable. Third, to better describe the multiscale properties of the features, we constructed a multiscale segmentation network to perform domain adaptation. Finally, to verify the performance of the proposed method, we conducted an experiment to compare the proposed method with three benchmark and four SOTA methods using three scene datasets. The mean intersection-over-union (mIOU) was significantly improved, and visualization results under multiple domain adaptation scenarios also show that the proposed method performed better than semantic segmentation methods.

PMID:39431138 | PMC:PMC11489487 | DOI:10.1016/j.fmre.2023.06.006

RSC Med Chem. 2024 Sep 26. doi: 10.1039/d4md00257a. Online ahead of print.

ABSTRACT

In light of the ongoing pandemic caused by SARS-CoV-2, effective and clinically translatable treatments are desperately needed for COVID-19 and its emerging variants. In this study, some derivatives, including 7β-aminocholestene compounds, and 3β-acetoxy-6-nitrocholesta-4,6-diene were synthesized, in quantitative yields from 7β-bromo-6-nitrocholest-5-enes (1-3) with a small library of amines. The synthesized steroidal products were then thoroughly characterized using a range of physicochemical techniques, including IR, NMR, UV, MS, and elemental analysis. Next, a virtual screening based on structures using docking studies was conducted to investigate the potential of these synthesized compounds as therapeutic candidates against SARS-CoV-2. Specifically, we evaluated the compounds' binding energy of the reactants and their products with three SARS-CoV-2 functional proteins: the papain-like protease, 3C-like protease or main protease, and RNA-dependent RNA polymerase. Our results indicate that the 7β-aminocholestene derivatives (4-8) display intermediate to excellent binding energy, suggesting that they interact strongly with the receptor's active amino acids and may be promising drug candidates for inhibiting SARS-CoV-2. Although the starting steroid derivatives; 7β-bromo-6-nitrocholest-5-enes (1-3) and one steroid product; 3β-acetoxy-6-nitrocholesta-4,6-diene (9) exhibited strong binding energies with various SARS-CoV-2 receptors, they did not meet the Lipinski Rule and ADMET properties required for drug development. These compounds showed either mutagenic or reproductive/developmental toxicity when assessed using toxicity prediction software. The findings based on structure-based virtual screening, suggest that 7β-aminocholestaines (4-8) may be useful for reducing the susceptibility to SARS-CoV-2 infection. The docking pose of compound 4, which has a high score of -7.4 kcal mol-1, was subjected to AI-assisted deep learning to generate 60 AI-designed molecules for drug design. Molecular docking of these AI molecules was performed to select optimal candidates for further analysis and visualization. The cytotoxicity and antioxidant effects of 3β-acetoxy-6-nitrocholesta-4,6-diene were tested in vitro, showing marked cytotoxicity and antioxidant activity. To elucidate the molecular basis for these effects, steroidal compound 9 was subjected to molecular docking analysis to identify potential binding interactions. The stability of the top-ranked docking pose was subsequently assessed using molecular dynamics simulations.

PMID:39430952 | PMC:PMC11485945 | DOI:10.1039/d4md00257a

Front Plant Sci. 2024 Oct 4;15:1433552. doi: 10.3389/fpls.2024.1433552. eCollection 2024.

ABSTRACT

The timely and accurate prediction of maize (Zea mays L.) yields prior to harvest is critical for food security and agricultural policy development. Currently, many researchers are using machine learning and deep learning to predict maize yields in specific regions with high accuracy. However, existing methods typically have two limitations. One is that they ignore the extensive correlation in maize planting data, such as the association of maize yields between adjacent planting locations and the combined effect of meteorological features and maize traits on maize yields. The other issue is that the performance of existing models may suffer significantly when some data in maize planting records is missing, or the samples are unbalanced. Therefore, this paper proposes an end-to-end bipartite graph neural network-based model for trait data imputation and yield prediction. The maize planting data is initially converted to a bipartite graph data structure. Then, a yield prediction model based on a bipartite graph neural network is developed to impute missing trait data and predict maize yield. This model can mine correlations between different samples of data, correlations between different meteorological features and traits, and correlations between different traits. Finally, to address the issue of unbalanced sample size at each planting location, we propose a loss function based on the gradient balancing mechanism that effectively reduces the impact of data imbalance on the prediction model. When compared to other data imputation and prediction models, our method achieves the best yield prediction result even when missing data is not pre-processed.

PMID:39430895 | PMC:PMC11486736 | DOI:10.3389/fpls.2024.1433552

NPJ Digit Med. 2024 Oct 20;7(1):294. doi: 10.1038/s41746-024-01292-5.

ABSTRACT

Alzheimer's disease (AD) is a global healthcare challenge lacking a simple and affordable detection method. We propose a novel deep learning framework, Eye-AD, to detect Early-onset Alzheimer's Disease (EOAD) and Mild Cognitive Impairment (MCI) using OCTA images of retinal microvasculature and choriocapillaris. Eye-AD employs a multilevel graph representation to analyze intra- and inter-instance relationships in retinal layers. Using 5751 OCTA images from 1671 participants in a multi-center study, our model demonstrated superior performance in EOAD (internal data: AUC = 0.9355, external data: AUC = 0.9007) and MCI detection (internal data: AUC = 0.8630, external data: AUC = 0.8037). Furthermore, we explored the associations between retinal structural biomarkers in OCTA images and EOAD/MCI, and the results align well with the conclusions drawn from our deep learning interpretability analysis. Our findings provide further evidence that retinal OCTA imaging, coupled with artificial intelligence, will serve as a rapid, noninvasive, and affordable dementia detection.

PMID:39428420 | DOI:10.1038/s41746-024-01292-5

Plant Methods. 2024 Oct 20;20(1):161. doi: 10.1186/s13007-024-01286-0.

ABSTRACT

BACKGROUND: The cotton whitefly (Bemisia tabaci) is a major global pest, causing significant crop damage through viral infestation and feeding. Traditional B. tabaci recognition relies on human eyes, which requires a large amount of work and high labor costs. The pests overlapping generations, high reproductive capacity, small size, and migratory behavior present challenges for the real-time monitoring and early warning systems. This study aims to develop an efficient, high-throughput automated system for detection of the cotton whiteflies. In this work, a novel tool for cotton whitefly fast identification and quantification was developed based on deep learning-based model. This approach enhances the effectiveness of B. tabaci control by facilitating earlier detection of its establishment in cotton, thereby allowing for a quicker implementation of management strategies.

RESULTS: We compiled a dataset of 1200 annotated images of whiteflies on cotton leaves, augmented using techniques like flipping and rotation. We modified the YOLO v8s model by replacing the C2f module with the Swin-Transformer and introducing a P2 structure in the Head, achieving a precision of 0.87, mAP50 of 0.92, and F1 score of 0.88 through ablation studies. Additionally, we employed SAHI for image preprocessing and integrated the whitefly detection algorithm on a Raspberry Pi, and developed a GUI-based visual interface. Our preliminary analysis revealed a higher density of whiteflies on cotton leaves in the afternoon and the middle-top, middle, and middle-down plant sections.

CONCLUSION: Utilizing the enhanced YOLO v8s deep learning model, we have achieved precise detection and counting of whiteflies, enabling its application on hardware devices like the Raspberry Pi. This approach is highly suitable for research requiring accurate quantification of cotton whiteflies, including phenotypic analyses. Future work will focus on deploying such equipment in large fields to manage whitefly infestations.

PMID:39427195 | DOI:10.1186/s13007-024-01286-0

BMC Pulm Med. 2024 Oct 19;24(1):523. doi: 10.1186/s12890-024-03344-8.

ABSTRACT

BACKGROUND: The quantitative analysis of computed tomography (CT) and Krebs von den Lungen-6 (KL-6) serum level has gained importance in the diagnosis, monitoring, and prognostication of interstitial lung disease (ILD). However, the associations between quantitative analysis of CT and serum KL-6 level remain poorly understood.

METHODS: In this retrospective observational study conducted at tertiary hospital between June 2020 and March 2022, quantitative analysis of CT was performed using the deep learning-based method including reticulation, ground glass opacity (GGO), honeycombing, and consolidation. We investigated the associations between CT-based phenotypes and serum KL-6 measured within three months of the CT scan. Furthermore, we evaluated the performance of the combined CT-based phenotypes and KL-6 levels in predicting hospitalizations due to respiratory reasons of ILD patients.

RESULTS: A total of 131 ILD patients (104 males) with a median age of 67 years were included in this study. Reticulation, GGO, honeycombing, and consolidation extents showed a positive correlation with KL-6 levels. [Reticulation, correlation coefficient (r) = 0.567, p < 0.001; GGO, r = 0.355, p < 0.001; honeycombing, r = 0.174, p = 0.046; and consolidation, r = 0.446, p < 0.001]. Additionally, the area under the ROC of the combined reticulation and KL-6 for hospitalizations due to respiratory reasons was 0.810 (p < 0.001).

CONCLUSIONS: Quantitative analysis of CT features and serum KL-6 levels ascertained a positive correlation between the two. In addition, the combination of reticulation and KL-6 shows potential for predicting hospitalizations of ILD patients due to respiratory causes. The combination of reticulation, focusing on phenotypic change in lung parenchyma, and KL-6, as an indicator of lung injury extent, could be helpful for monitoring and predicting the prognosis of various types of ILD.

PMID:39427156 | DOI:10.1186/s12890-024-03344-8

BMC Infect Dis. 2024 Oct 19;24(1):1181. doi: 10.1186/s12879-024-10032-9.

ABSTRACT

Lumpy skin disease virus (LSDV) is an extremely infectious, viral, and chronic skin disease that is caused by the Capripox virus. This viral disease is predominantly found in cows. Mosquitoes and ticks are the primary transmitters for the spread of this virus. Recently, LSDV has been rapidly spreading all over the world, especially in several areas of Pakistan, India, and Iran. Thousands of cows have died due to this infectious virus in Pakistan and early detection of LSDV is needed to avoid further loss. The prediction and classification of LSDV are hindered by the lack of publicly available datasets. Despite a few studies using LSDV datasets, such datasets are often small, which may lead to model overfitting. In this regard, we collect the dataset from several online sources, as well as, collecting images from veterinary farms in different areas of Pakistan. Deep learning has been widely used in the medical field for disease detection and classification. Therefore, this study leverages DenseNet deep learning models for LSDV detection and classification. Experiments are performed using VGG-16, ResNet-50, MobileNet-V2, custom-designed convolutional neural network, and Inception-V3. The DenseNet architecture presents a Convolutional Block Attention Module (CBAM) and Spatial Attention (SA) for the prediction and classification of LSD. Results demonstrate that a 99.11% accuracy can be obtained on the augmented dataset while a 94.23% accuracy can be achieved with the original dataset for chicken pox, monkey pox, and LSDV. Comparison with state-of-the-art studies corroborates the superior performance of the proposed model.

PMID:39427155 | DOI:10.1186/s12879-024-10032-9

BMC Musculoskelet Disord. 2024 Oct 19;25(1):822. doi: 10.1186/s12891-024-07934-9.

ABSTRACT

The accurate classification of bone tumours is crucial for guiding clinical decisions regarding treatment and follow-up. However, differentiating between various tumour types is challenging due to the rarity of certain entities, high intra-class variability, and limited training data in clinical practice. This study proposes a multimodal deep learning model that integrates clinical metadata and X-ray imaging to improve the classification of primary bone tumours. The dataset comprises 1,785 radiographs from 804 patients collected between 2000 and 2020, including metadata such as age, affected bone site, tumour position, and gender. Ten tumour types were selected, with histopathology or tumour board decisions serving as the reference standard.

METHODS: Our model is based on the NesT image classification model and a multilayer perceptron with a joint fusion architecture. Descriptive statistics included incidence and percentage ratios for discrete parameters, and mean, standard deviation, median, and interquartile range for continuous parameters.

RESULTS: The mean age of the patients was 33.62 ± 18.60 years, with 54.73% being male. Our multimodal deep learning model achieved 69.7% accuracy in classifying primary bone tumours, outperforming the Vision Transformer model by five percentage points. SHAP values indicated that age had the most substantial influence among the considered metadata.

CONCLUSION: The joint fusion approach developed in this study, integrating clinical metadata and imaging data, outperformed state-of-the-art models in classifying primary bone tumours. The use of SHAP values provided insights into the impact of different metadata on the model's performance, highlighting the significant role of age. This approach has potential implications for improving diagnostic accuracy and understanding the influence of clinical factors in tumour classification.

PMID:39427131 | DOI:10.1186/s12891-024-07934-9

NPJ Digit Med. 2024 Oct 20;7(1):293. doi: 10.1038/s41746-024-01290-7.

ABSTRACT

The success of deep learning (DL) relies heavily on training data from which DL models encapsulate information. Consequently, the development and deployment of DL models expose data to potential privacy breaches, which are particularly critical in data-sensitive contexts like medicine. We propose a new technique named DiffGuard that generates realistic and diverse synthetic medical images with annotations, even indistinguishable for experts, to replace real data for DL model training, which cuts off their direct connection and enhances privacy safety. We demonstrate that DiffGuard enhances privacy safety with much less data leakage and better resistance against privacy attacks on data and model. It also improves the accuracy and generalizability of DL models for segmentation and classification of mediastinal neoplasms in multi-center evaluation. We expect that our solution would enlighten the road to privacy-preserving DL for precision medicine, promote data and model sharing, and inspire more innovation on artificial-intelligence-generated-content technologies for medicine.

PMID:39427092 | DOI:10.1038/s41746-024-01290-7

Sci Rep. 2024 Oct 19;14(1):24567. doi: 10.1038/s41598-024-76047-z.

ABSTRACT

Survival analysis in medical research has witnessed a growing interest in applying deep learning techniques to model complex, high-dimensional, heterogeneous, incomplete, and censored data. Current methods make assumptions about the relations between data that may not be valid in practice. Therefore, we introduce SAVAE (Survival Analysis Variational Autoencoder). SAVAE, based on Variational Autoencoders, contributes significantly to the field by introducing a tailored Evidence Lower BOund formulation, supporting various parametric distributions for covariates and survival time (if the log-likelihood is differentiable). It offers a general method that demonstrates robustness and stability through different experiments. Our proposal effectively estimates time-to-event, accounting for censoring, covariate interactions, and time-varying risk associations. We validate our model in diverse datasets, including genomic, clinical, and demographic tabular data, with varying levels of censoring. This approach demonstrates competitive performance compared to state-of-the-art techniques, as assessed by the Concordance Index and the Integrated Brier Score. SAVAE also offers an interpretable model that parametrically models covariates and time. Moreover, its generative architecture facilitates further applications such as clustering, data imputation, and synthetic patient data generation through latent space inference from survival data. This approach fosters data sharing and collaboration, improving medical research and personalized patient care.

PMID:39427084 | DOI:10.1038/s41598-024-76047-z

Sci Rep. 2024 Oct 19;14(1):24624. doi: 10.1038/s41598-024-75826-y.

ABSTRACT

Various machine learning (ML) and deep learning (DL) techniques have been recently applied to the forecasting of laboratory earthquakes from friction experiments. The magnitude and timing of shear failures in stick-slip cycles are predicted using features extracted from the recorded ultrasonic or acoustic emission (AE) signals. In addition, the Rate and State Friction (RSF) constitutive laws are extensively used to model the frictional behavior of faults. In this work, we use data from shear experiments coupled with passive acoustic (variance, kurtosis, and AE rate) interleaved with active source ultrasonic monitoring (transmitted wave amplitude) to develop physics-informed neural network (PINN) models incorporating the RSF law and AE rate generation equation with wave amplitude serving as a proxy for friction state variable. This PINN framework allows learning RSF parameters from stick-slip experiments rather than measuring them through a series of velocity step experiments. We observe that when the stick-slip cycles are irregular, the PINN models outperform the data-driven DL models. Transfer learning (TL) PINN models are also developed by pre-training on data collected at one normal stress level followed by forecasting shear failures and retrieving RSF parameters at other stress levels (i.e., with different recurrence intervals) after retraining on a limited amount of new data. Our findings suggest that TL models perform better compared to standalone models. Both standalone and TL PINN-estimated RSF parameters and their ground truth values show excellent agreements thus demonstrating that RSF parameters can be retrieved from laboratory stick-slip experiments using the corresponding acoustic data and that the transmitted wave amplitude provides a good representation of the evolving frictional state during stick-slips.

PMID:39427066 | DOI:10.1038/s41598-024-75826-y

NPJ Precis Oncol. 2024 Oct 19;8(1):235. doi: 10.1038/s41698-024-00710-x.

ABSTRACT

Deep learning models for predicting variant pathogenicity have not been thoroughly evaluated on real-world clinical phenotypes. Here, we apply state-of-the-art pathogenicity prediction models to hereditary breast cancer gene variants in UK Biobank participants. Model predictions for missense variants in BRCA1, BRCA2 and PALB2, but not ATM and CHEK2, were associated with breast cancer risk. However, deep learning models had limited clinical utility when specifically applied to variants of uncertain significance.

PMID:39427061 | DOI:10.1038/s41698-024-00710-x

Sci Rep. 2024 Oct 19;14(1):24564. doi: 10.1038/s41598-024-75385-2.

ABSTRACT

This paper introduces a novel application of spatial-temporal graph neural networks (ST-GNNs) to predict groundwater levels. Groundwater level prediction is inherently complex, influenced by various hydrological, meteorological, and anthropogenic factors. Traditional prediction models often struggle with the nonlinearity and non-stationary characteristics of groundwater data. Our study leverages the capabilities of ST-GNNs to address these challenges in the Overbetuwe area, Netherlands. We utilize a comprehensive dataset encompassing 395 groundwater level time series and auxiliary data such as precipitation, evaporation, river stages, and pumping well data. The graph-based framework of our ST-GNN model facilitates the integration of spatial interconnectivity and temporal dynamics, capturing the complex interactions within the groundwater system. Our modified Multivariate Time Graph Neural Network model shows significant improvements over traditional methods, particularly in handling missing data and forecasting future groundwater levels with minimal bias. The model's performance is rigorously evaluated when trained and applied with both synthetic and measured data, demonstrating superior accuracy and robustness in comparison to traditional numerical models in long-term forecasting. The study's findings highlight the potential of ST-GNNs in environmental modeling, offering a significant step forward in predictive modeling of groundwater levels.

PMID:39427045 | DOI:10.1038/s41598-024-75385-2