Radiat Oncol. 2024 Nov 25;19(1):170. doi: 10.1186/s13014-024-02531-5.

ABSTRACT

BACKGROUND: Accurate calculation of lung cancer dose using the Monte Carlo (MC) algorithm in CyberKnife (CK) is essential for precise planning. We aim to employ deep learning to directly predict the 3D dose distribution calculated by the MC algorithm, enabling rapid and accurate automatic planning. However, most current methods solely focus on conventional intensity-modulated radiation therapy and assume a consistent beam configuration across all patients. This study seeks to develop a more versatile model incorporating variable beam configurations of CK and considering the patient's anatomy.

METHODS: This study proposed that the AB (anatomy and beam) model be compared with the control Mask (only anatomy) model. These models are based on a 3D U-Net network to investigate the impact of CK beam encoding information on dose prediction. The study collected 86 lung cancer patients who received CK's built-in MC algorithm plans using different beam configurations for training/validation (66 cases) and testing (20 cases). We compared the gamma passing rate, dose difference maps, and relevant dose-volume metrics to evaluate the model's performance. In addition, the Dice similarity coefficients (DSCs) were calculated to assess the spatial correspondence of isodose volumes.

RESULTS: The AB model demonstrated superior performance compared to the Mask model, particularly in the trajectory dose of the beam. The DSCs of the AB model were 20-40% higher than that of the Mask model in some dose regions. We achieved approximately 99% for the PTV and generally more than 95% for the organs at risk (OARs) referred to the clinical planning dose in the gamma passing rates (3 mm/3%). Relative to the Mask model, the AB model exhibited more than 90% improvement in small voxels (p < 0.001). The AB model matched well with the clinical plan's dose-volume histograms, and the average dose error for all organs was 1.65 ± 0.69%.

CONCLUSIONS: Our proposed new model signifies a crucial advancement in predicting CK 3D dose distributions for clinical applications. It enables planners to rapidly and precisely predict MC doses for lung cancer based on patient-specific beam configurations and optimize the CK treatment process.

PMID:39587661 | DOI:10.1186/s13014-024-02531-5

J Cheminform. 2024 Nov 25;16(1):130. doi: 10.1186/s13321-024-00921-1.

ABSTRACT

Non-Contact Atomic Force Microscopy with CO-functionalized metal tips (referred to as HR-AFM) provides access to the internal structure of individual molecules adsorbed on a surface with totally unprecedented resolution. Previous works have shown that deep learning (DL) models can retrieve the chemical and structural information encoded in a 3D stack of constant-height HR-AFM images, leading to molecular identification. In this work, we overcome their limitations by using a well-established description of the molecular structure in terms of topological fingerprints, the 1024-bit Extended Connectivity Chemical Fingerprints of radius 2 (ECFP4), that were developed for substructure and similarity searching. ECFPs provide local structural information of the molecule, each bit correlating with a particular substructure within the molecule. Our DL model is able to extract this optimized structural descriptor from the 3D HR-AFM stacks and use it, through virtual screening, to identify molecules from their predicted ECFP4 with a retrieval accuracy on theoretical images of 95.4%. Furthermore, this approach, unlike previous DL models, assigns a confidence score, the Tanimoto similarity, to each of the candidate molecules, thus providing information on the reliability of the identification. By construction, the number of times a certain substructure is present in the molecule is lost during the hashing process, necessary to make them useful for machine learning applications. We show that it is possible to complement the fingerprint-based virtual screening with global information provided by another DL model that predicts from the same HR-AFM stacks the chemical formula, boosting the identification accuracy up to a 97.6%. Finally, we perform a limited test with experimental images, obtaining promising results towards the application of this pipeline under real conditions.Scientific contributionPrevious works on molecular identification from AFM images used chemical descriptors that were intuitive for humans but sub-optimal for neural networks. We propose a novel method to extract the ECFP4 from AFM images and identify the molecule via a virtual screening, beating previous state-of-the-art models.

PMID:39587659 | DOI:10.1186/s13321-024-00921-1

BMC Public Health. 2024 Nov 25;24(1):3269. doi: 10.1186/s12889-024-20713-4.

ABSTRACT

BACKGROUND: Workplace may not only increase the risk of heat-related illnesses and injuries but also compromise work efficiency, particularly in a warming climate. This study aimed to utilize machine learning (ML) and deep learning (DL) algorithms to quantify the impact of temperature discomfort on productivity loss among petrochemical workers and to identify key influencing factors.

METHODS: A cross-sectional face-to-face questionnaire survey was conducted among petrochemical workers between May and September 2023 in Fujian Province, China. Initial feature selection was performed using Lasso regression. The dataset was divided into training (70%), validation (20%), and testing (10%) sets. Six predictive models were evaluated: support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), Gaussian Naive Bayes (GNB), multilayer perceptron (MLP), and logistic regression (LR). The most effective model was further analyzed with SHapley Additive exPlanations (SHAP).

RESULTS: Among the 2393 workers surveyed, 58.4% (1,747) reported productivity loss when working in high temperatures. Lasso regression identified twenty-seven predictive factors such as educational level and smoking. All six models displayed strong prediction accuracy (SVM = 0.775, RF = 0.760, XGBoost = 0.727, GNB = 0.863, MLP = 0.738, LR = 0.680). GNB model showed the best performance, with a cutoff of 0.869, accuracy of 0.863, precision of 0.897, sensitivity of 0.918, specificity of 0.715, and an F1-score of 0.642, indicating its efficacy as a predictive tool. SHAP analysis showed that occupational health training (SHAP value: -3.56), protective measures (-2.61), and less physically demanding jobs (-1.75) were negatively associated with heat-attributed productivity loss, whereas lack of air conditioning (1.92), noise (2.64), vibration (1.15), and dust (0.95) increased the risk of heat-induced productivity loss.

CONCLUSIONS: Temperature discomfort significantly undermined labor productivity in the petrochemical sector, and this impact may worsen in a warming climate if adaptation and prevention measures are insufficient. To effectively reduce heat-related productivity loss, there is a need to strengthen occupational health training and implement strict controls for occupational hazards, minimizing the potential combined effects of heat with other exposures.

PMID:39587532 | DOI:10.1186/s12889-024-20713-4

BMC Pregnancy Childbirth. 2024 Nov 25;24(1):777. doi: 10.1186/s12884-024-06862-9.

ABSTRACT

BACKGROUND: Changes in body temperature anticipate labor onset in numerous mammals, yet this concept has not been explored in humans. We investigated if continuous body temperature exhibits similar changes in women and whether these changes may be linked to hormonal status. Finally, we developed a deep learning model using temperature patterning to provide a daily forecast of time to labor onset.

METHODS: We evaluated patterns in continuous skin temperature data in 91 (n = 54 spontaneous labors) pregnant women using a wearable smart ring. In a subset of 28 pregnancies, we examined daily steroid hormone samples leading up to labor to analyze relationships among hormones and body temperature trajectory. Finally, we applied an autoencoder long short-term memory (AE-LSTM) deep learning model to provide a novel daily estimation of days until labor onset.

RESULTS: Features of temperature change leading up to labor were associated with urinary hormones and labor type. Spontaneous labors exhibited greater estriol to α-pregnanediol ratio, as well as lower body temperature and more stable circadian rhythms compared to pregnancies that did not undergo spontaneous labor. Skin temperature data from 54 pregnancies that underwent spontaneous labor between 34 and 42 weeks of gestation were included in training the AE-LSTM model, and an additional 37 pregnancies that underwent artificial induction of labor or Cesarean without labor were used for further testing. The input to the pipeline was 5-min skin temperature data from a gestational age of 240 days until the day of labor onset. During cross-validation AE-LSTM average error (true - predicted) dropped below 2 days at 8 days before labor, independent of gestational age. Labor onset windows were calculated from the AE-LSTM output using a probabilistic distribution of model error. For these windows AE-LSTM correctly predicted labor start for 79% of the spontaneous labors within a 4.6-day window at 7 days before true labor, and 7.4-day window at 10 days before true labor.

CONCLUSION: Continuous skin temperature reflects progression toward labor and hormonal change during pregnancy. Deep learning using continuous temperature may provide clinically valuable tools for pregnancy care.

PMID:39587525 | DOI:10.1186/s12884-024-06862-9

Expert Rev Med Devices. 2024 Nov 25. doi: 10.1080/17434440.2024.2434732. Online ahead of print.

ABSTRACT

INTRODUCTION: Diagnosing oral cancer is crucial in healthcare, with technological advancements enhancing early detection and outcomes. This review examines the impact of handheld AI-based tools, focusing on Convolutional Neural Networks (CNNs) and their advanced architectures in oral cancer diagnosis.

METHODS: A comprehensive search across PubMed, Scopus, Google Scholar, and Web of Science identified papers on deep learning (DL) in oral cancer diagnosis using digital images. The review, registered with PROSPERO, employed PRISMA and QUADAS-2 for search and risk assessment, with data analyzed through bubble and bar charts.

RESULTS: Twenty-five papers were reviewed, highlighting classification, segmentation, and object detection as key areas. Despite challenges like limited annotated datasets and data imbalance, models such as DenseNet121, VGG19, and EfficientNet-B0 excelled in binary classification, while EfficientNet-B4, Inception-V4, and Faster R-CNN were effective for multiclass classification and object detection. Models achieved up to 100% precision, 99% specificity, and 97.5% accuracy, showcasing AI's potential to improve diagnostic accuracy. Combining datasets and leveraging transfer learning enhances detection, particularly in resource-limited settings.

CONCLUSION: Handheld AI tools are transforming oral cancer diagnosis, with ethical considerations guiding their integration into healthcare systems. DL offers explainability, builds trust in AI-driven diagnoses, and facilitates telemedicine integration.

PMID:39587051 | DOI:10.1080/17434440.2024.2434732

Radiol Med. 2024 Nov 26. doi: 10.1007/s11547-024-01909-5. Online ahead of print.

ABSTRACT

PURPOSE: To develop and validate deep learning (DL) models using preoperative contrast-enhanced CT images for tumor auto-segmentation and microsatellite instability (MSI) prediction in colorectal cancer (CRC).

MATERIALS AND METHODS: Patients with CRC who underwent surgery or biopsy between January 2018 and April 2023 were retrospectively enrolled. Mismatch repair protein expression was determined via immunohistochemistry or fluorescence multiplex polymerase chain reaction-capillary electrophoresis. Manually delineated tumor contours using arterial and venous phase CT images by three abdominal radiologists are served as ground truth. Tumor auto-segmentation used nnU-Net. MSI prediction employed ViT or convolutional neural networks models, trained and validated with arterial and venous phase images (image model) or combined clinical-pathological factors (combined model). The segmentation model was evaluated using patch coverage ratio, Dice coefficient, recall, precision, and F1-score. The predictive models' efficacy was assessed using areas under the curves and decision curve analysis.

RESULTS: Overall, 2180 patients (median age: 61 years ± 17 [SD]; 1285 males) were divided into training (n = 1159), validation (n = 289), and independent external test (n = 732) groups. High-level MSI status was present in 435 patients (20%). In the external test set, the segmentation model performed well in the arterial phase, with patch coverage ratio, Dice coefficient, recall, precision, and F1-score values of 0.87, 0.71, 0.72, 0.74, and 0.71, respectively. For MSI prediction, the combined models outperformed the clinical model (AUC = 0.83 and 0.82 vs 0.67, p < 0.001) and two image models (AUC = 0.75 and 0.77, p < 0.001). Decision curve analysis confirmed the higher net benefit of the combined model compared to the other models across probability thresholds ranging from 0.1 to 0.45.

CONCLUSION: DL enhances tumor segmentation efficiency and, when integrated with contrast-enhanced CT and clinicopathological factors, exhibits good diagnostic performance in predicting MSI in CRC.

PMID:39586941 | DOI:10.1007/s11547-024-01909-5

J Imaging Inform Med. 2024 Nov 25. doi: 10.1007/s10278-024-01315-3. Online ahead of print.

ABSTRACT

Accurate classification of cells in bronchoalveolar lavage (BAL) fluid is essential for the assessment of lung disease in pneumology and critical care medicine. However, the effectiveness of BAL fluid analysis is highly dependent on individual expertise. Our research is focused on improving the accuracy and efficiency of BAL cell classification using the "You Only Look Once" (YOLO) algorithm to reduce variability and increase the accuracy of cell detection in BALF analysis. We assess various YOLOv7 iterations, including YOLOv7, YOLOv7 with Adam and label smoothing, YOLOv7-E6E, and YOLOv7-E6E with Adam and label smoothing focusing on the detection of four key cell types of diagnostic importance in BAL fluid: macrophages, lymphocytes, neutrophils, and eosinophils. This study utilised cytospin preparations of BAL fluid, employing May-Grunwald-Giemsa staining, and analysed a dataset comprising 2032 images with 42,221 annotations. Classification performance was evaluated using recall, precision, F1 score, mAP@.5, and mAP@.5;.95 along with a confusion matrix. The comparison of four algorithmic approaches revealed minor distinctions in mean results, falling short of statistical significance (p < 0.01; p < 0.05). YOLOv7, with an inference time of 13.5 ms for 640 × 640 px images, achieved commendable performance across all cell types, boasting an average F1 metric of 0.922, precision of 0.916, recall of 0.928, and mAP@.5 of 0.966. Remarkably, all four cell types were classified consistently with high-performance metrics. Notably, YOLOv7 demonstrated marginally superior class value dispersion when compared to YOLOv7-adam-label-smoothing, YOLOv7-E6E, and YOLOv7-E6E-adam-label-smoothing, albeit without statistical significance. Consequently, there is limited justification for deploying the more computationally intensive YOLOv7-E6E and YOLOv7-E6E-adam-label-smoothing models. This investigation indicates that the default YOLOv7 variant is the preferred choice for differential cytology due to its accessibility, lower computational demands, and overall more consistent results than more complex models.

PMID:39586912 | DOI:10.1007/s10278-024-01315-3

J Imaging Inform Med. 2024 Nov 25. doi: 10.1007/s10278-024-01340-2. Online ahead of print.

ABSTRACT

To investigate whether automatic segmentation based on DCE-MRI with a deep learning (DL) algorithm enabled advantages over manual segmentation in differentiating BI-RADS 4 breast lesions. A total of 197 patients with suspicious breast lesions from two medical centers were enrolled in this study. Patients treated at the First Hospital of Qinhuangdao between January 2018 and April 2024 were included as the training set (n = 138). Patients treated at Lanzhou University Second Hospital were assigned to an external validation set (n = 59). Areas of suspicious lesions were delineated based on DL automatic segmentation and manual segmentation, and evaluated consistency through the Dice correlation coefficient. Radiomics models were constructed based on DL and manual segmentations to predict the nature of BI-RADS 4 lesions. Meanwhile, the nature of the lesions was evaluated by both a professional radiologist and a non-professional radiologist. Finally, the area under the curve value (AUC) and accuracy (ACC) were used to determine which prediction model was more effective. Sixty-four malignant cases (32.5%) and 133 benign cases (67.5%) were included in this study. The DL-based automatic segmentation model showed high consistency with manual segmentation, achieving a Dice coefficient of 0.84 ± 0.11. The DL-based radiomics model demonstrated superior predictive performance compared to professional radiologists, with an AUC of 0.85 (95% CI 0.79-0.92). The DL model significantly reduced working time and improved efficiency by 83.2% compared to manual segmentation, further demonstrating its feasibility for clinical applications. The DL-based radiomics model for automatic segmentation outperformed professional radiologists in distinguishing between benign and malignant lesions in BI-RADS category 4, thereby helping to avoid unnecessary biopsies. This groundbreaking progress suggests that the DL model is expected to be widely applied in clinical practice in the near future, providing an effective auxiliary tool for the diagnosis and treatment of breast cancer.

PMID:39586911 | DOI:10.1007/s10278-024-01340-2

NPJ Precis Oncol. 2024 Nov 25;8(1):274. doi: 10.1038/s41698-024-00760-1.

ABSTRACT

Predicting H3K27M mutation status in midline gliomas non-invasively is of considerable interest, particularly using deep learning with 11C-methionine (MET) and 18F-fluoroethyltyrosine (FET) positron emission tomography (PET). To optimise prediction efficiency, we derived an assistance training (AT) scheme to allow mutual benefits between MET and FET learning to boost the predictability but still only require either PET as inputs for predictions. Our method significantly surpassed conventional convolutional neural network (CNN), radiomics-based, and MR-based methods, achieved an area under the curve (AUC) of 0.9343 for MET, and an AUC of 0.8619 for FET during internal cross-validation (n = 90). The performance remained high in hold-out testing (n = 19) and consecutive testing cohorts (n = 21), with AUCs of 0.9205 and 0.7404. The clinical feasibility of the proposed method was confirmed by the agreements to multi-departmental decisions and outcomes in pathology-uncertain cases. The findings positions our method as a promising tool for aiding treatment decisions in midline glioma.

PMID:39587279 | DOI:10.1038/s41698-024-00760-1

Sci Rep. 2024 Nov 25;14(1):29241. doi: 10.1038/s41598-024-80292-7.

ABSTRACT

The aim of this study was to develop a quantitative feature-based model from histopathologic images to assess the prognosis of patients with gastric cancer. Whole slide image (WSI) images of H&E-stained histologic specimens of gastric cancer patients from The Cancer Genome Atlas were included and randomly assigned to training and test groups in a 7:3 ratio. A systematic preprocessing approach was employed as well as a non-overlapping segmentation method that combined patch-level prediction with a multi-instance learning approach to integrate features across the slide images. Subjects were categorized into high- or low-risk groups based on the median risk score derived from the model, and the significance of this stratification was assessed using a log-rank test. In addition, combining transcriptomic data from patients and data from other large cohort studies, we further searched for genes associated with pathological features and their prognostic value. A total of 165 gastric cancer patients were included for model training, and a total of 26 features were integrated through multi-instance learning, with each process generating 11 probabilistic features and 2 predictive labeling features. We applied a 10-fold Lasso-Cox regression model to achieve dimensionality reduction of these features. The predictive accuracy of the model was verified using Kaplan-Meyer (KM) curves for stratification with a consistency index of 0.741 for the training set and 0.585 for the test set. Deep learning-based resultant supervised pathohistological features have the potential for superior prognostic stratification of gastric cancer patients, transforming image pixels into an effective and labor-saving tool to optimize the clinical management of gastric cancer patients. Also, SLITRK4 was identified as a prognostic marker for gastric cancer.

PMID:39587240 | DOI:10.1038/s41598-024-80292-7

Sci Rep. 2024 Nov 25;14(1):29141. doi: 10.1038/s41598-024-80940-y.

ABSTRACT

Efficiently predicting the paratope holds immense potential for enhancing antibody design, treating cancers and other serious diseases, and advancing personalized medicine. Although traditional methods are highly accurate, they are often time-consuming, labor-intensive, and reliant on 3D structures, restricting their broader use. On the other hand, machine learning-based methods, besides relying on structural data, entail descriptor computation, consideration of diverse physicochemical properties, and feature engineering. Here, we develop a deep learning-assisted prediction method for paratope identification, relying solely on amino acid sequences and being antigen-agnostic. Built on the ProtTrans architecture, and utilizing pre-trained protein and antibody language models, we extract efficient embeddings for predicting paratope. By incorporating positional encoding for Complementarity Determining Regions, our model gains a deeper structural understanding, achieving remarkable performance with a 0.904 ROC AUC, 0.701 F1-score, and 0.585 MCC on benchmark datasets. In addition to yielding accurate antibody paratope predictions, our method exhibits strong performance in predicting nanobody paratope, achieving a ROC AUC of 0.912 and a PR AUC of 0.665 on the nanobody dataset. Notably, our approach outperforms structure-based prediction methods, boasting a PR AUC of 0.731. Various conducted ablation studies, which elaborate on the impact of each part of the model on the prediction task, show that the improvement in prediction performance by applying CDR positional encoding together with CNNs depends on the specific protein and antibody language models used. These results highlight the potential of our method to advance disease understanding and aid in the discovery of new diagnostics and antibody therapies.

PMID:39587231 | DOI:10.1038/s41598-024-80940-y

Sci Rep. 2024 Nov 26;14(1):29263. doi: 10.1038/s41598-024-80548-2.

ABSTRACT

Alzheimer's disease (AD), a prevalent neurodegenerative disorder, leads to progressive dementia, which impairs decision-making, problem-solving, and communication. While there is no cure, early detection can facilitate treatments to slow its progression. Deep learning (DL) significantly enhances AD detection by analyzing brain imaging data to identify early biomarkers, improving diagnostic accuracy and predicting disease progression more precisely than traditional methods. In this article, we propose an ensemble methodology for DL models to detect AD from brain MRIs. We trained an enhanced Xception architecture once to produce multiple snapshots, providing diverse insights into MRI features. A decision-level fusion strategy was employed, combining decision scores with a RF meta-learner using a blending algorithm. The efficacy of our ensemble technique is confirmed by the experimental findings, which categorize Alzheimer's into four groups with 99.14% accuracy. This methodology may help medical practitioners provide patients with Alzheimer's with individualized care. Subsequent efforts will concentrate on enhancing the model's efficacy via its generalization to a variety of datasets.

PMID:39587224 | DOI:10.1038/s41598-024-80548-2

Sci Rep. 2024 Nov 25;14(1):29129. doi: 10.1038/s41598-024-80366-6.

ABSTRACT

In this work, the dynamic mechanical properties of concrete‒granite composites with various roughness interfaces were investigated via split Hopkinson pressure bar (SHPB) system to evaluate the impact resistance of the lining‒surrounding rock composite structure that is commonly present in rock engineering. The dynamic uniaxial compressive strength of the composite at an impact speed of 11.3 m/s may increase by 20.55% when the joint roughness coefficient (JRC) increases from 0 to 28.64 according to the experimental results. The JRC increased the strain rate effect of the composite but reduced the confining pressure effect. The relationships between the dynamic deformation parameters, such as the elastic modulus and critical strain, and the study variables, as well as the correlation mechanism between the macroscopic and microscopic failure morphologies of the composite and the stress‒strain curves, were investigated. A developed long short-term memory (LSTM) deep learning method was utilized to predict the dynamic stress‒strain relationships of the composites after 144 sets of experimental data were split into training and testing sets. The prediction issue of peak stress for the composites after varying time steps was presented to the recursive neural network LSTM, which was evaluated and compared with the traditional back propagation neural network (BPNN) and random forest (RF). The LSTM model showed the strongest prediction capacity when considering accuracy and predictive evaluation indicators. Compared with the BPNN model, the RF model was worse at capturing the viscoelastic plastic properties of the constitutive model while having superior assessment indicators.

PMID:39587222 | DOI:10.1038/s41598-024-80366-6

Sci Rep. 2024 Nov 25;14(1):29159. doi: 10.1038/s41598-024-79001-1.

ABSTRACT

Ensuring robust network security is crucial in the context of Software-Defined Networking(SDN). Which, becomes a multi-billion dollar industry, and it's deployed in many data centers nowadays. The new technology provides network programmability, network centralized control, and a global view of the network. But, unfortunately, it comes with new vulnerabilities, and new attack vectors compared to the traditional network. SDN network cybersecurity became a trending research topic due to the hype of Machine Learning (ML) when a group of Machine Learning(ML) techniques called Deep Learning(DL) started to take shape in the setting of SDN networks. This paper focuses on developing advanced Deep Learning(DL) models to address the inherent new attack vectors. In this paper, we have built and compared two models that can be used for building a complete Intrusion Detection System(IDS) solution, one using a hybrid CNN-LSTM architecture and the other using Transformer encoder-only architecture. We specifically target the SDN controller where it represents a crucial point. We utilized the InSDN dataset for training and testing our models, this dataset captures real-world traffic within the SDN environment. For evaluation, we have used accuracy, precision, recall, and F1 Score. Our experiment results show that the Transformer model with 48 features achieves the highest accuracy at 99.02%, while the CNN-LSTM model achieves 99.01%. We have reduced the features to 6 and 4, which gave us varying impacts on the models' performance. We have merged 4 poorly represented attacks in one class, which enhanced the accuracy by a significant score. Additionally, we investigate binary classification by merging all attack types into a single class, as a result, the accuracy increased for both models. The CNN-LSTM model achieves the best results with an accuracy of 99.19% for 6 feature sets, this enhances the state-of-the-art results.

PMID:39587182 | DOI:10.1038/s41598-024-79001-1

Sci Rep. 2024 Nov 25;14(1):29249. doi: 10.1038/s41598-024-79348-5.

ABSTRACT

This bicenter retrospective analysis included 162 patients who had undergone prostate biopsy following prebiopsy MRI, excluding those with PCa identified only in the peripheral zone (PZ). DLR T2WI achieved a 69% reduction in scan time relative to TSE T2WI. The intermethod agreement between the two T2WI sets in terms of the Prostate Imaging Reporting and Data System (PI-RADS) classification and extraprostatic extension (EPE) grade was measured using the intraclass correlation coefficient (ICC) and diagnostic performance was assessed with the area under the receiver operating characteristic curve (AUC). Clinically significant PCa (csPCa) was found in 74 (45.7%) patients. Both T2WI methods showed high intermethod agreement for the overall PI-RADS classification (ICC: 0.907-0.949), EPE assessment (ICC: 0.925-0.957) and lesion size measurement (ICC: 0.980-0.996). DLR T2WI and TSE T2WI showed similar AUCs (0.666-0.814 versus 0.684-0.832) for predicting EPE. The AUCs for detecting csPCa with DLR T2WI (0.834-0.935) and TSE T2WI (0.891-0.935) were comparable in 139 patients with TZ lesions with no significant differences (P > 0.05). The findings suggest that DLR T2WI is an efficient alternative for TZ lesion assessment, offering reduced scan times without compromising diagnostic accuracy.

PMID:39587164 | DOI:10.1038/s41598-024-79348-5

Sci Rep. 2024 Nov 25;14(1):29199. doi: 10.1038/s41598-024-79621-7.

ABSTRACT

Operator learning is a rising field of scientific computing where inputs or outputs of a machine learning model are functions defined in infinite-dimensional spaces. In this paper, we introduce Neon (Neural Epistemic Operator Networks), an architecture for generating predictions with uncertainty using a single operator network backbone, which presents orders of magnitude less trainable parameters than deep ensembles of comparable performance. We showcase the utility of this method for sequential decision-making by examining the problem of composite Bayesian Optimization (BO), where we aim to optimize a function f = g ∘ h , where h : X → C ( Y , R d s ) is an unknown map which outputs elements of a function space, and g : C ( Y , R d s ) → R is a known and cheap-to-compute functional. By comparing our approach to other state-of-the-art methods on toy and real world scenarios, we demonstrate that Neon achieves state-of-the-art performance while requiring orders of magnitude less trainable parameters.

PMID:39587149 | DOI:10.1038/s41598-024-79621-7

Sci Rep. 2024 Nov 25;14(1):29224. doi: 10.1038/s41598-024-79838-6.

ABSTRACT

While deep learning (DL) offers the compelling ability to detect details beyond human vision, its black-box nature makes it prone to misinterpretation. A key problem is algorithmic shortcutting, where DL models inform their predictions with patterns in the data that are easy to detect algorithmically but potentially misleading. Shortcutting makes it trivial to create models with surprisingly accurate predictions that lack all face validity. This case study shows how easily shortcut learning happens, its danger, how complex it can be, and how hard it is to counter. We use simple ResNet18 convolutional neural networks (CNN) to train models to do two things they should not be able to do: predict which patients avoid consuming refried beans or beer purely by examining their knee X-rays (AUC of 0.63 for refried beans and 0.73 for beer). We then show how these models' abilities are tied to several confounding and latent variables in the image. Moreover, the image features the models use to shortcut cannot merely be removed or adjusted through pre-processing. The end result is that we must raise the threshold for evaluating research using CNNs to proclaim new medical attributes that are present in medical images.

PMID:39587148 | DOI:10.1038/s41598-024-79838-6

Sci Data. 2024 Nov 25;11(1):1283. doi: 10.1038/s41597-024-04159-2.

ABSTRACT

MedSegBench is a comprehensive benchmark designed to evaluate deep learning models for medical image segmentation across a wide range of modalities. It covers a wide range of modalities, including 35 datasets with over 60,000 images from ultrasound, MRI, and X-ray. The benchmark addresses challenges in medical imaging by providing standardized datasets with train/validation/test splits, considering variability in image quality and dataset imbalances. The benchmark supports binary and multi-class segmentation tasks with up to 19 classes and uses the U-Net architecture with various encoder/decoder networks such as ResNets, EfficientNet, and DenseNet for evaluations. MedSegBench is a valuable resource for developing robust and flexible segmentation algorithms and allows for fair comparisons across different models, promoting the development of universal models for medical tasks. It is the most comprehensive study among medical segmentation datasets. The datasets and source code are publicly available, encouraging further research and development in medical image analysis.

PMID:39587124 | DOI:10.1038/s41597-024-04159-2

J Phys Chem B. 2024 Nov 25. doi: 10.1021/acs.jpcb.4c02507. Online ahead of print.

ABSTRACT

Intrinsically disordered Proteins (IDPs) constitute a large and structureless class of proteins with significant functions. The existence of IDPs challenges the conventional notion that the biological functions of proteins rely on their three-dimensional structures. Despite lacking well-defined spatial arrangements, they exhibit diverse biological functions, influencing cellular processes and shedding light on disease mechanisms. However, it is expensive to run experiments or simulations to characterize this class of proteins. Consequently, we designed an ML model that relies solely on amino acid sequences. In this study, we introduce the IDP-Bert model, a deep-learning architecture leveraging Transformers and Protein Language Models to map sequences directly to IDP properties. Our experiments demonstrate accurate predictions of IDP properties, including Radius of Gyration, end-to-end Decorrelation Time, and Heat Capacity.

PMID:39586094 | DOI:10.1021/acs.jpcb.4c02507

PLoS One. 2024 Nov 25;19(11):e0313323. doi: 10.1371/journal.pone.0313323. eCollection 2024.

ABSTRACT

Sea turtles exhibit high migratory rates and occupy a broad range of habitats, which in turn makes monitoring these taxa challenging. Applying deep learning (DL) models to vast image datasets collected from citizen science programs can offer promising solutions to overcome the challenge of monitoring the wide habitats of wildlife, particularly sea turtles. Among DL models, object detection models, such as the You Only Look Once (YOLO) series, have been extensively employed for wildlife classification. Despite their successful application in this domain, detecting objects in images with complex backgrounds, including underwater environments, remains a significant challenge. Recently, instance segmentation models have been developed to address this issue by providing more accurate classification of complex images compared to traditional object detection models. This study compared the performance of two state-of-the-art DL methods namely; the object detection model (YOLOv5) and instance segmentation model (YOLOv5-seg), to detect and classify sea turtles. The images were collected from iNaturalist and Google and then divided into 64% for training, 16% for validation, and 20% for test sets. Model performance during and after finishing training was evaluated by loss functions and various indexes, respectively. Based on loss functions, YOLOv5-seg demonstrated a lower error rate in detecting rather than classifying sea turtles than the YOLOv5. According to mean Average Precision (mAP) values, which reflect precision and recall, the YOLOv5-seg model showed superior performance than YOLOv5. The mAP0.5 and mAP0.5:0.95 for the YOLOv5 model were 0.885 and 0.795, respectively, whereas for the YOLOv5-seg, these values were 0.918 and 0.831, respectively. In particular, based on the loss functions and classification results, the YOLOv5-seg showed improved performance for detecting rather than classifying sea turtles compared to the YOLOv5. The results of this study may help improve sea turtle monitoring in the future.

PMID:39585892 | DOI:10.1371/journal.pone.0313323