Cancer Res Commun. 2024 Dec 31. doi: 10.1158/2767-9764.CRC-24-0397. Online ahead of print.

ABSTRACT

Intratumor heterogeneity (ITH) presents challenges for precision oncology, but methods for its spatial quantification, scalable at population levels, do not exist. Based on previous work showing that admixture of PAM50 subtype can be measured from bulk tissue using transcriptomic data, we trained a deep neural network (DNN) to quantify subtype ITH in Luminal A (LumA) breast cancer from routinely-stained whole slide images. We tested the hypothesis that subtype admixture detected in images was associated with tumor aggressiveness and adverse outcome. In 680 cases from the TCGA-BRCA cohort, we determined adherence to assigned subtype by applying matrix factorization to each transcriptome. The purest cases were split into groups for initial testing, training and parameter tuning. 230 LumA-assigned cases were held out for final testing. Image patches were fed into a DNN pre-trained on histology images. We measured the association of tumor area classified as LumA in the image to tumor characteristics and survival. Among LumA-assigned cases, admixture was associated with slightly higher ER-positivity but lower PR-positivity and ER-related gene expression, and higher HER2-positivity, tumor size, grade, and TNM stage. Image admixture was associated with more TP53 and less PIK3CA mutation. Progression-free survival was significantly shorter in more admixed cases. Our findings demonstrate that deep learning, trained to recognize genomic correlates in tissue morphology, can quantify and map subtype admixture in LumA breast cancer that has clinical significance. The low-cost and scalability of this method holds potential as a research tool for investigating ITH and perhaps improving the efficacy of precision oncology.

PMID:39740059 | DOI:10.1158/2767-9764.CRC-24-0397

PLoS One. 2024 Dec 31;19(12):e0310925. doi: 10.1371/journal.pone.0310925. eCollection 2024.

ABSTRACT

This study aimed to evaluate the performance of a deep learning-based segmentation model for predicting outcomes of non-surgical endodontic treatment. Preoperative and 3-year postoperative periapical radiographic images of each tooth from routine root canal treatments performed by endodontists from 2015 to 2021 were obtained retrospectively from Thammasat University hospital. Preoperative radiographic images of 1200 teeth with 3-year follow-up results (440 healed, 400 healing, and 360 disease) were collected. Mask Region-based Convolutional Neural Network (Mask R-CNN) was used to pixel-wise segment the root from other structures in the image and trained to predict class label into healed, healing and disease. Three endodontists annotated 1080 images used for model training, validation, and testing. The performance of the model was evaluated on a test set and also by comparison with the performance of clinicians (general practitioners and endodontists) with and without the help of the model on independent 120 images. The performance of the Mask R-CNN prediction model was high with the mean average precision (mAP) of 0.88 (95% CI 0.83-0.93) and area under the precision-recall curve of 0.91 (95% CI 0.88-0.94), 0.83 (95% CI 0.81-0.85), 0.91 (95% CI 0.90-0.92) on healed, healing and disease, respectively. The prediction metrics of general practitioners and endodontists significantly improved with the help of Mask R-CNN outperforming clinicians alone with mAP increasing from 0.75 (95% CI 0.72-0.78) to 0.84 (95% CI 0.81-0.87) and 0.88 (95% CI 0.85-0.91) to 0.92 (95% CI 0.89-0.95), respectively. In conclusion, deep learning-based segmentation model had the potential to predict non-surgical endodontic treatment outcomes from periapical radiographic images and were expected to aid in endodontic treatment.

PMID:39739891 | DOI:10.1371/journal.pone.0310925

BMC Neurosci. 2024 Dec 31;25(1):80. doi: 10.1186/s12868-024-00913-9.

ABSTRACT

Obstructive sleep apnea (OSA) is widespread, under-recognized, and under-treated, impacting the health and quality of life for millions. The current gold standard for sleep apnea testing is based on the in-lab sleep study, which is costly, cumbersome, not readily available and represents a well-known roadblock to managing this huge societal burden. Assessment of neuromuscular function involved in the upper airway using electromyography (EMG) has shown potential to characterize and diagnose sleep apnea, while the development of transmembranous electromyography (tmEMG), a painless surface probe, has made this opportunity practical and highly feasible. However, experience and ability to interpret electrical signals from the upper airway are scarce, and much of the pertinent information within the signal is likely difficult to detect visually. To overcome this issue, we explored the use of transformers, a deep learning (DL) model architecture with attention mechanisms, to model tmEMG data and distinguish between electromyographic signals from a cohort of control, neurogenic, and sleep apnea patients. Our approach involved three strategies to train a generalizable model on a relatively small dataset including, (1) transfer learning using an audio spectral transformer (AST), (2) the use of 6,000 simulated EMG recordings, converted to spectrograms and using standard backpropagation for fine-tuning, and (3) application of regularization to prevent overfitting and enhance generalizability. This DL approach was tested using 177 transoral EMG recordings from a prior study's database that included six healthy controls, five moderate to severe OSA patients, and five amyotrophic lateral sclerosis (ALS) patients with evidence of bulbar involvement (neurogenic injury). Sensitivity and specificity for classifying neurogenic cases from controls were 98% and 73%, respectively, while classifying OSA from controls were 88% and 64%, respectively. Notably, by averaging the predicted probabilities of each segment for individual patients, the model correctly classified up to 82% of control and OSA patients. These results not only suggest a potential to diagnose OSA patients accurately, but also to identify OSA endotypes that involve neuromuscular pathology, which has major implications for clinical management, patient outcomes, and research.

PMID:39741274 | DOI:10.1186/s12868-024-00913-9

BMC Bioinformatics. 2024 Dec 31;25(1):393. doi: 10.1186/s12859-024-06024-w.

ABSTRACT

BACKGROUND: The collection of substantial amounts of electroencephalogram (EEG) data is typically time-consuming and labor-intensive, which adversely impacts the development of decoding models with strong generalizability, particularly when the available data is limited. Utilizing sufficient EEG data from other subjects to aid in modeling the target subject presents a potential solution, commonly referred to as domain adaptation. Most current domain adaptation techniques for EEG decoding primarily focus on learning shared feature representations through domain alignment strategies. Since the domain shift cannot be completely removed, target EEG samples located near the edge of clusters are also susceptible to misclassification.

METHODS: We propose a novel adaptive deep feature representation (ADFR) framework to improve the cross-subject EEG classification performance through learning transferable EEG feature representations. Specifically, we first minimize the distribution discrepancy between the source and target domains by employing maximum mean discrepancy (MMD) regularization, which aids in learning the shared feature representations. We then utilize the instance-based discriminative feature learning (IDFL) regularization to make the learned feature representations more discriminative. Finally, the entropy minimization (EM) regularization is further integrated to adjust the classifier to pass through the low-density region between clusters. The synergistic learning between above regularizations during the training process enhances EEG decoding performance across subjects.

RESULTS: The effectiveness of the ADFR framework was evaluated on two public motor imagery (MI)-based EEG datasets: BCI Competition III dataset 4a and BCI Competition IV dataset 2a. In terms of average accuracy, ADFR achieved improvements of 3.0% and 2.1%, respectively, over the state-of-the-art methods on these datasets.

CONCLUSIONS: The promising results highlight the effectiveness of the ADFR algorithm for EEG decoding and show its potential for practical applications.

PMID:39741250 | DOI:10.1186/s12859-024-06024-w

Nat Methods. 2024 Dec 31. doi: 10.1038/s41592-024-02557-3. Online ahead of print.

ABSTRACT

Optical approaches to monitor neural activity are transforming neuroscience, owing to a fast-evolving palette of genetically encoded molecular reporters. However, the field still requires robust and label-free technologies to monitor the multifaceted biomolecular changes accompanying brain development, aging or disease. Here, we have developed vibrational fiber photometry as a low-invasive method for label-free monitoring of the biomolecular content of arbitrarily deep regions of the mouse brain in vivo through spontaneous Raman spectroscopy. Using a tapered fiber probe as thin as 1 µm at its tip, we elucidate the cytoarchitecture of the mouse brain, monitor molecular alterations caused by traumatic brain injury, as well as detect markers of brain metastasis with high accuracy. We view our approach, which introduces a deep learning algorithm to suppress probe background, as a promising complement to the existing palette of tools for the optical interrogation of neural function, with application to fundamental and preclinical investigations of the brain and other organs.

PMID:39741190 | DOI:10.1038/s41592-024-02557-3

Sci Rep. 2024 Dec 31;14(1):32169. doi: 10.1038/s41598-024-83517-x.

ABSTRACT

In the last decade, grating-based phase-contrast computed tomography (gbPC-CT) has received growing interest. It provides additional information about the refractive index decrement in the sample. This signal shows an increased soft-tissue contrast. However, the resolution dependence of the signal poses a challenge: its contrast enhancement is overcompensated by the low resolution in low-dose applications such as clinical computed tomography. As a result, the implementation of gbPC-CT is currently tied to a higher dose. To reduce the dose, we introduce the self-supervised deep learning network Noise2Inverse into the field of gbPC-CT. We evaluate the behavior of the Noise2Inverse parameters on the phase-contrast results. Afterward, we compare its results with other denoising methods, namely the Statistical Iterative Reconstruction, Block Matching 3D, and Patchwise Phase Retrieval. In the example of Noise2Inverse, we show that deep learning networks can deliver superior denoising results with respect to the investigated image quality metrics. Their application allows to increase the resolution while maintaining the dose. At higher resolutions, gbPC-CT can naturally deliver higher contrast than conventional absorption-based CT. Therefore, the application of machine learning-based denoisers shifts the dose-normalized image quality in favor of gbPC-CT, bringing it one step closer to medical application.

PMID:39741166 | DOI:10.1038/s41598-024-83517-x

Commun Biol. 2024 Dec 31;7(1):1718. doi: 10.1038/s42003-024-07415-8.

ABSTRACT

The functional organization of the human object vision pathway distinguishes between animate and inanimate objects. To understand animacy perception, we explore the case of zoomorphic objects resembling animals. While the perception of these objects as animal-like seems obvious to humans, such "Animal bias" is a striking discrepancy between the human brain and deep neural networks (DNNs). We computationally investigated the potential origins of this bias. We successfully induced this bias in DNNs trained explicitly with zoomorphic objects. Alternative training schedules failed to cause an Animal bias. We considered the superordinate distinction between animate and inanimate classes, the sensitivity for faces and bodies, the bias for shape over texture, the role of ecologically valid categories, recurrent connections, and language-informed visual processing. These findings provide computational support that the Animal bias for zoomorphic objects is a unique property of human perception yet can be explained by human learning history.

PMID:39741161 | DOI:10.1038/s42003-024-07415-8

Light Sci Appl. 2025 Jan 1;14(1):12. doi: 10.1038/s41377-024-01678-w.

ABSTRACT

Metamaterials have revolutionized wave control; in the last two decades, they evolved from passive devices via programmable devices to sensor-endowed self-adaptive devices realizing a user-specified functionality. Although deep-learning techniques play an increasingly important role in metamaterial inverse design, measurement post-processing and end-to-end optimization, their role is ultimately still limited to approximating specific mathematical relations; the metamaterial is still limited to serving as proxy of a human operator, realizing a predefined functionality. Here, we propose and experimentally prototype a paradigm shift toward a metamaterial agent (coined metaAgent) endowed with reasoning and cognitive capabilities enabling the autonomous planning and successful execution of diverse long-horizon tasks, including electromagnetic (EM) field manipulations and interactions with robots and humans. Leveraging recently released foundation models, metaAgent reasons in high-level natural language, acting upon diverse prompts from an evolving complex environment. Specifically, metaAgent's cerebrum performs high-level task planning in natural language via a multi-agent discussion mechanism, where agents are domain experts in sensing, planning, grounding, and coding. In response to live environmental feedback within a real-world setting emulating an ambient-assisted living context (including human requests in natural language), our metaAgent prototype self-organizes a hierarchy of EM manipulation tasks in conjunction with commanding a robot. metaAgent masters foundational EM manipulation skills related to wireless communications and sensing, and it memorizes and learns from past experience based on human feedback.

PMID:39741131 | DOI:10.1038/s41377-024-01678-w

Am J Crit Care. 2025 Jan 1;34(1):41-51. doi: 10.4037/ajcc2025597.

ABSTRACT

BACKGROUND: Hyperkalemia can be detected by point-of-care (POC) blood testing and by artificial intelligence- enabled electrocardiography (ECG). These 2 methods of detecting hyperkalemia have not been compared.

OBJECTIVE: To determine the accuracy of POC and ECG potassium measurements for hyperkalemia detection in patients with critical illness.

METHODS: This retrospective study involved intensive care patients in an academic medical center from October 2020 to September 2021. Patients who had 12-lead ECG, POC potassium measurement, and central laboratory potassium measurement within 1 hour were included. The POC potassium measurements were obtained from arterial blood gas analysis; ECG potassium measurements were calculated by a previously developed deep learning model. Hyperkalemia was defined as a central laboratory potassium measurement of 5.5 mEq/L or greater.

RESULTS: Fifteen patients with hyperkalemia and 252 patients without hyperkalemia were included. The POC and ECG potassium measurements were available about 35 minutes earlier than central laboratory results. Correlation with central laboratory potassium measurement was better for POC testing than for ECG (mean absolute errors of 0.211 mEq/L and 0.684 mEq/L, respectively). For POC potassium measurement, area under the receiver operating characteristic curve (AUC) to detect hyperkalemia was 0.933, sensitivity was 73.3%, and specificity was 98.4%. For ECG potassium measurement, AUC was 0.884, sensitivity was 93.3%, and specificity was 63.5%.

CONCLUSIONS: The ECG potassium measurement, with its high sensitivity and coverage rate, may be used initially and followed by POC potassium measurement for rapid detection of life-threatening hyperkalemia.

PMID:39740977 | DOI:10.4037/ajcc2025597

Brain Behav. 2025 Jan;15(1):e70206. doi: 10.1002/brb3.70206.

ABSTRACT

PURPOSE: A debilitating and poorly understood symptom of Parkinson's disease (PD) is freezing of gait (FoG), which increases the risk of falling. Clinical evaluations of FoG, relying on patients' subjective reports and manual examinations by specialists, are unreliable, and most detection methods are influenced by subject-specific factors.

METHOD: To address this, we developed a novel algorithm for detecting FoG events based on movement signals. To enhance efficiency, we propose a novel architecture integrating a bottleneck attention module into a standard bidirectional long short-term memory network (BiLSTM). This architecture, adaptable to a convolution bottleneck attention-BiLSTM (CBA-BiLSTM), classifies signals using data from ankle, leg, and trunk sensors.

FINDING: Given three movement directions from three locations, we reduce computational complexity in two phases: selecting optimal channels through ensemble learning followed by feature reduction using attention mapping. In FoG event detection tests, performance improved significantly compared to control groups and existing methods, achieving 99.88% accuracy with only two channels.

CONCLUSION: The reduced computational complexity enables real-time monitoring. Our approach demonstrates substantial improvements in classification results compared to traditional deep learning methods.

PMID:39740772 | DOI:10.1002/brb3.70206

Eur J Radiol. 2024 Dec 22;183:111899. doi: 10.1016/j.ejrad.2024.111899. Online ahead of print.

ABSTRACT

OBJECTIVE: To assess whether CT style conversion between different CT vendors using a routable generative adversarial network (RouteGAN) could minimize variation in ILD quantification, resulting in improved functional correlation of quantitative CT (QCT) measures.

METHODS: Patients with idiopathic pulmonary fibrosis (IPF) who underwent unenhanced chest CTs with vendor A and a pulmonary function test (PFT) were retrospectively evaluated. As deep-learning based ILD quantification software was mainly developed using vendor B CT, style-converted images from vendor A to B style were generated using RouteGAN. Quantification was performed in both original and converted images. Measurement variability in QCT between original and converted images was evaluated using the concordance correlation coefficient (CCC). Two radiologists visually evaluated quantification accuracy using original and converted images. Correlations between CT parameters and PFT measures were assessed.

RESULTS: Total 112 patients (mean age, 61; 82 men) were studied. Measurement variability between original and converted CT was a CCC of 0.20 for reticulation, 0.72 for honeycombing, and 0.59 for ground-glass opacity. The median visual accuracy scores were higher for the quantification using converted compared with the original images (P < 0.001). Correlation between fibrosis score increased significantly after CT conversion for both forced vital capacity (original vs. converted; -0.35 vs. -0.50; P = 0.005) and diffusing capacity of the lung for carbon monoxide (-0.50 vs. -0.66; P < 0.001).

CONCLUSION: The improved accuracy in deep learning based ILD quantification after applying GAN-based CT style conversion can result in the improved functional correlation of QCT measurements in patients with IPF.

PMID:39740598 | DOI:10.1016/j.ejrad.2024.111899

Am J Otolaryngol. 2024 Nov 30;46(1):104549. doi: 10.1016/j.amjoto.2024.104549. Online ahead of print.

ABSTRACT

BACKGROUND: Hyperostosis is a common radiographic feature of inverted papilloma (IP) tumor origin on computed tomography (CT). Herein, we developed a machine learning (ML) model capable of analyzing CT images and identifying IP attachment sites.

METHODS: A retrospective review of patients treated for IP at our institution was performed. The tumor attachment site was manually segmented on CT scans by the operating surgeon. We used a nnU-Net model, a state-of-the-art deep learning-based segmentation algorithm that automatically configures image preprocessing, network architecture, training, and post-processing to identify the IP attachment site. The model was trained and evaluated using a 5-fold cross validation, where each iteration split the data into train/validation/test to avoid chances of overfitting. The attachment site was classified as either 'identified or 'not identified' using the nnU-Net model output and the Sørensen-Dice coefficient (Dice) was used to further evaluate the segmentation performance of each subject.

RESULTS: A total of 58 subjects met enrollment criteria. The algorithm identified the attachment site in 55.2 % (n = 32) of patients with an average dice score (+/-SD) of 0.34 (+/- 0.24). In the univariate analysis, the algorithm performed better for attachment sites within the maxillary sinus (OR 4.0; p < 0.05) and performed worse during revision surgery (OR 0.13; p < 0.05). Multivariate logistic regression analysis confirmed these associations for maxillary attachment site (OR 4.6; p < 0.05) and revision surgery (OR 0.11; p < 0.05).

CONCLUSION: A state-of-the-art ML model successfully identified the attachment site of IP with a high degree of fidelity in select cases, but requires larger sample sizes and more diverse datasets to become reliably integrated into clinical practice.

PMID:39740533 | DOI:10.1016/j.amjoto.2024.104549

Sci Rep. 2024 Dec 30;14(1):31813. doi: 10.1038/s41598-024-82956-w.

ABSTRACT

Analyzing microbial samples remains computationally challenging due to their diversity and complexity. The lack of robust de novo protein function prediction methods exacerbates the difficulty in deriving functional insights from these samples. Traditional prediction methods, dependent on homology and sequence similarity, often fail to predict functions for novel proteins and proteins without known homologs. Moreover, most of these methods have been trained on largely eukaryotic data, and have not been evaluated on or applied to microbial datasets. This research introduces DeepGOMeta, a deep learning model designed for protein function prediction as Gene Ontology (GO) terms, trained on a dataset relevant to microbes. The model is applied to diverse microbial datasets to demonstrate its use for gaining biological insights. Data and code are available at https://github.com/bio-ontology-research-group/deepgometa.

PMID:39738309 | DOI:10.1038/s41598-024-82956-w

Sci Rep. 2024 Dec 30;14(1):31759. doi: 10.1038/s41598-024-82022-5.

ABSTRACT

Accurate classification of logos is a challenging task in image recognition due to variations in logo size, orientation, and background complexity. Deep learning models, such as VGG16, have demonstrated promising results in handling such tasks. However, their performance is highly dependent on optimal hyperparameter settings, whose fine-tuning is both labor-intensive and time-consuming. Swarm intelligence algorithms have been widely adopted to solve many highly nonlinear, multimodal problems and have succeeded significantly. The Hunger Games Search (HGS) is a recent swarm intelligence algorithm that has shown good performance across various applications. However, the standard HGS still faces limitations, such as restricted population diversity and a tendency to get trapped in local optima, which can hinder its effectiveness. In this paper, we propose an optimized deep learning architecture called EHGS-VGG16 designed based on the VGG16 model and boosted by an enhanced Hunger Games Search (EHGS) algorithm for hyperparameter tuning. The proposed enhancement to HGS involves modified search strategies, incorporating the concepts of "local best" and a "local escaping mechanism" to improve its exploration capability. To validate our approach, the evaluation is conducted in three folds. First, the EHGS algorithm is evaluated through 30 real-valued benchmark functions from the IEEE CEC2014 suite. Second, a custom-developed VGG16 model is tested on the Flickr-27 logo classification dataset and compared against state-of-the-art deep learning models such as ResNet50V2, InceptionV3, DenseNet121, EfficientNetB0, and MobileNetV2. Finally, EHGS is integrated into the VGG16 model to optimize its hyperparameters. The experimental results show that VGG16 outperformed the other counterparts with an accuracy of 0.956966, a precision of 0.957137, and a recall of 0.956966. Moreover, the integration of EHGS further improved classification quality by 3%. These findings highlight the potential of combining evolutionary optimization techniques with deep learning for enhanced accuracy in log classification tasks.

PMID:39738231 | DOI:10.1038/s41598-024-82022-5

Sci Rep. 2024 Dec 30;14(1):31700. doi: 10.1038/s41598-024-81724-0.

ABSTRACT

Accurate diagnosis of oral lesions, early indicators of oral cancer, is a complex clinical challenge. Recent advances in deep learning have demonstrated potential in supporting clinical decisions. This paper introduces a deep learning model for classifying oral lesions, focusing on accuracy, interpretability, and reducing dataset bias. The model integrates three components: (i) a Classification Stream, utilizing a CNN to categorize images into 16 lesion types (baseline model), (ii) a Guidance Stream, which aligns class activation maps with clinically relevant areas using ground truth segmentation masks (GAIN model), and (iii) an Anatomical Site Prediction Stream, improving interpretability by predicting lesion location (GAIN+ASP model). The development dataset comprised 2765 intra-oral digital images of 16 lesion types from 1079 patients seen at an oral pathology clinic between 1999 and 2021. The GAIN model demonstrated a 7.2% relative improvement in accuracy over the baseline for 16-class classification, with superior class-specific balanced accuracy and AUC scores. Additionally, the GAIN model enhanced lesion localization and improved the alignment between attention maps and ground truth. The proposed models also exhibited greater robustness against dataset bias, as shown in ablation studies.

PMID:39738228 | DOI:10.1038/s41598-024-81724-0

Sci Rep. 2024 Dec 30;14(1):31624. doi: 10.1038/s41598-024-78148-1.

ABSTRACT

Deep learning models are widely used for traffic forecasting on freeways due to their ability to learn complex temporal and spatial relationships. In particular, graph neural networks, which integrate graph theory into deep learning, have become popular for modeling traffic sensor networks. However, traditional graph convolutional networks (GCNs) face limitations in capturing long-range spatial correlations, which can hinder accurate long-term predictions. To address this issue, we propose the Two-level Resolution Neural Network, which enhances interpretability by introducing two resolution blocks. The first block captures large-scale regional traffic patterns, while the second block, using a GCN, focuses on small-scale spatial correlations, informed by the regional predictions. This structure allows the model to intuitively integrate both local and distant traffic data, improving long-term forecasting. In addition to its predictive capabilities, TwoResNet offers enhanced interpretability, particularly in scenarios involving noisy or incomplete data.

PMID:39738225 | DOI:10.1038/s41598-024-78148-1

Sci Rep. 2024 Dec 30;14(1):31668. doi: 10.1038/s41598-024-81188-2.

ABSTRACT

Immune checkpoint inhibitor (ICI) treatment has proven successful for advanced melanoma, but is associated with potentially severe toxicity and high costs. Accurate biomarkers for response are lacking. The present work is the first to investigate the value of deep learning on CT imaging of metastatic lesions for predicting ICI treatment outcomes in advanced melanoma. Adult patients that were treated with ICI for advanced melanoma were retrospectively identified from ten participating centers. A deep learning model (DLM) was trained on volumes of lesions on baseline CT to predict clinical benefit. The DLM was compared to and combined with a model of known clinical predictors (presence of liver and brain metastasis, level of lactate dehydrogenase, performance status and number of affected organs). A total of 730 eligible patients with 2722 lesions were included. The DLM reached an area under the receiver operating characteristic (AUROC) of 0.607 [95%CI 0.565-0.648]. In comparison, a model of clinical predictors reached an AUROC of 0.635 [95%CI 0.59 -0.678]. The combination model reached an AUROC of 0.635 [95% CI 0.595-0.676]. Differences in AUROC were not statistically significant. The output of the DLM was significantly correlated with four of the five input variables of the clinical model. The DLM reached a statistically significant discriminative value, but was unable to improve over known clinical predictors. The present work shows that the assessment over known clinical predictors is an essential step for imaging-based prediction and brings important nuance to the almost exclusively positive findings in this field.

PMID:39738216 | DOI:10.1038/s41598-024-81188-2

Sci Rep. 2024 Dec 30;14(1):31667. doi: 10.1038/s41598-024-81228-x.

ABSTRACT

Nursing activity recognition has immense importance in the development of smart healthcare management and is an extremely challenging area of research in human activity recognition. The main reasons are an extreme class-imbalance problem and intra-class variability depending on both the subject and the recipient. In this paper, we apply a unique two-step feature extraction, coupled with an intermediate feature 'Angle' and a new feature called mean min max sum to render the features robust against intra-class variation. After intermediate and final feature extraction, we use an ensemble of a random forest classifier and a stacked convolutional neural network (S-CNN) model to detect activities and users. Unlike traditional CNN, the S-CNN takes the input feature channels in separate pathways with equal importance, which makes it robust to intra-class variation and produces accurate results. We apply this method to two benchmark open-source nurse care activity data sets. Our algorithm is robust enough to recognize both activity and user (Nurse) simultaneously. During the recognition process, this algorithm automatically finds the important features in the data set. Using this algorithm, the highest testing accuracies were achieved for activity recognition on the two (publicly available in IEEE DataPort) benchmark data sets: The CARECOM Nurse Care Activity (70.6% accuracy) and the Heiseikai Nurse Care Activity data set (85.7% accuracy). Moreover, the highest accuracy achieved for user identification on Data Set 1 and Data Set 2 is 78.2% and 92.7%, respectively.

PMID:39738208 | DOI:10.1038/s41598-024-81228-x

Sci Rep. 2024 Dec 30;14(1):31564. doi: 10.1038/s41598-024-68920-8.

ABSTRACT

Prediction and discovery of new materials with desired properties are at the forefront of quantum science and technology research. A major bottleneck in this field is the computational resources and time complexity related to finding new materials from ab initio calculations. In this work, an effective and robust deep learning-based model is proposed by incorporating persistent homology with graph neural network which offers an accuracy of 91.4 % and an F1 score of 88.5 % in classifying topological versus non-topological materials, outperforming the other state-of-the-art classifier models. Additionally, out-of-distribution and newly discovered topological materials can be classified using our method with high confidence. The incorporation of the graph neural network encodes the underlying relation between the atoms into the model based on their crystalline structures and thus proved to be an effective method to represent and process non-Euclidean data like molecules with a relatively shallow network. The persistent homology pipeline in the proposed neural network integrates a topological analysis of crystal structures into the deep learning model, enhancing both robustness and performance. Our classifier can serve as an efficacious tool for predicting the topological class, thereby enabling a high-throughput search for fascinating topological materials.

PMID:39738190 | DOI:10.1038/s41598-024-68920-8

Sci Rep. 2024 Dec 30;14(1):31603. doi: 10.1038/s41598-024-78865-7.

ABSTRACT

Manual segmentation of lesions, required for radiotherapy planning and follow-up, is time-consuming and error-prone. Automatic detection and segmentation can assist radiologists in these tasks. This work explores the automated detection and segmentation of brain metastases (BMs) in longitudinal MRIs. It focuses on several important aspects: identifying and segmenting new lesions for screening and treatment planning, re-segmenting lesions in successive images using prior lesion locations as an additional input channel, and performing multi-component segmentation to distinguish between enhancing tissue, edema, and necrosis. The key component of the proposed approach is to propagate the lesion mask from the previous time point to improve the detection performance, which we refer to as "re-segmentation". The retrospective data includes 518 metastases in 184 contrast-enhanced T1-weighted MRIs originating from 49 patients (63% male, 37% female). 131 time-points (36 patients, 418 BMs) are used for cross-validation, the remaining 53 time-points (13 patients, 100 BMs) are used for testing. The lesions were manually delineated with label 1: enhancing lesion, label 2: edema, and label 3: necrosis. One-tailed t-tests are used to compare model performance including multiple segmentation and detection metrics. Significance is considered as p < 0.05. A Dice Similarity Coefficient (DSC) of 0.79 and F 1 -score of 0.80 are obtained for the segmentation of new lesions. On follow-up scans, the re-segmentation model significantly outperforms the segmentation model (DSC and F 1 0.78 and 0.88 vs 0.56 and 0.60). The re-segmentation model also significantly outperforms the simple segmentation model on the enhancing lesion (DSC 0.76 vs 0.53) and edema (0.52 vs 0.47) components, while similar scores are obtained on the necrosis component (0.62 vs 0.63). Additionally, we analyze the correlation between lesion size and segmentation performance, as demonstrated in various studies that highlight the challenges in segmenting small lesions. Our findings indicate that this correlation disappears when utilizing the re-segmentation approach and evaluating with the unbiased normalized DSC. In conclusion, the automated segmentation of new lesions and subsequent re-segmentation in follow-up images was achievable, with high level of performance obtained for single- and multiple-component segmentation tasks.

PMID:39738168 | DOI:10.1038/s41598-024-78865-7