Sci Rep. 2024 Nov 26;14(1):29390. doi: 10.1038/s41598-024-79011-z.

ABSTRACT

Recommendation Systems (RS) play a crucial role in delivering personalized item suggestions, yet traditional methods often struggle with accuracy, scalability, efficiency, and cold-start challenges. This paper presents the HRS-IU-DL model, a novel hybrid recommendation system that advances the field by integrating multiple sophisticated techniques to enhance both accuracy and relevance. The proposed model uniquely combines user-based and item-based Collaborative Filtering (CF) to effectively analyze user-item interactions, Neural Collaborative Filtering (NCF) to capture complex non-linear relationships, and Recurrent Neural Networks (RNN) to identify sequential patterns in user behavior. Furthermore, it incorporates Content-Based Filtering (CBF) with Term Frequency-Inverse Document Frequency (TF-IDF) for in-depth analysis of item attributes. A key contribution of this work is the innovative fusion of CF, NCF, RNN, and CBF, which collectively address significant challenges such as data sparsity, the cold-start problem, and the increasing demand for personalized recommendations. Additionally, the model employs N-Sample techniques to recommend the top 10 similar items based on user-specified genres, leveraging methods like Cosine Similarity, Singular Value Decomposition (SVD), and TF-IDF. The HRS-IU-DL model is rigorously evaluated on the publicly available Movielens 100k dataset using train-test splits. Performance is assessed using metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Precision, and Recall. The results demonstrate that the HRS-IU-DL model not only outperforms state-of-the-art approaches but also achieves substantial improvements across these evaluation metrics, highlighting its contribution to the advancement of RS technology.

PMID:39592677 | DOI:10.1038/s41598-024-79011-z

Sci Rep. 2024 Nov 26;14(1):29382. doi: 10.1038/s41598-024-79361-8.

ABSTRACT

Echocardiography is the gold standard of diagnosis and evaluation of patent ductus arteriosus (PDA), a common condition among preterm infants that can cause hemodynamic abnormalities and increased mortality rates, but this technique requires a skilled specialist and is not always available. Meanwhile, chest X-ray (CXR) imaging is also known to exhibit signs of PDA and is a routine imaging modality in neonatal intensive care units. In this study, we aim to find and objectively define CXR image features that are associated with PDA by training and visually analyzing a deep learning model. We first collected 4617 echocardiograms from neonatal intensive care unit patients and 17,448 CXR images that were taken 4 days before to 3 days after the echocardiograms were obtained. We trained a deep learning model to predict the presence of severe PDA using the CXR images, and then visualized the model using GradCAM++ to identify the regions of the CXR images important for the model's prediction. The visualization results showed that the model focused on the regions around the upper thorax, lower left heart, and lower right lung. Based on these results, we hypothesized and evaluated three radiographic features of PDA: cardiothoracic ratio, upper heart width to maximum heart width ratio, and upper heart width to thorax width ratio. We then trained an XGBoost model to predict the presence of severe PDA using these radiographic features combined with clinical features. The model achieved an AUC of 0.74, with a high specificity of 0.94. Our study suggests that the proposed radiographic features of CXR images can be used as an auxiliary tool to predict the presence of PDA in preterm infants. This can be useful for the early detection of PDA in neonatal intensive care units in cases where echocardiography is not available.

PMID:39592675 | DOI:10.1038/s41598-024-79361-8

Sci Rep. 2024 Nov 26;14(1):29285. doi: 10.1038/s41598-024-79632-4.

ABSTRACT

With the fast-growing interconnection of smart technologies, the Industrial Internet of Things (IIoT) has revolutionized how industries work by connecting devices and sensors and automating regular operations via the Internet of Things (IoTs). IoT devices provide seamless diversity and connectivity in different application domains. This system and its transmission channels are subjected to targeted cyberattacks due to their round-the-clock connectivity. Accordingly, a multilevel security solution is needed to safeguard the industrial system. By analyzing the data packet, the Intrusion Detection System (IDS) counteracts the cyberattack for the targeted attack in the IIoT platform. Various research has been undertaken to address the concerns of cyberattacks on IIoT networks using machine learning (ML) and deep learning (DL) approaches. This study introduces a new Bayesian Machine Learning with the Sparrow Search Algorithm for Cyberattack Detection (BMLSSA-CAD) technique in the IIoT networks. The proposed BMLSSA-CAD technique aims to enhance security in IIoT networks by detecting cyberattacks. In the BMLSSA-CAD technique, the min-max scaler normalizes the input dataset. Additionally, the method utilizes the Chameleon Optimization Algorithm (COA)-based feature selection (FS) approach to identify the optimal feature set. The BMLSSA-CAD technique uses the Bayesian Belief Network (BBN) model for cyberattack detection. The hyperparameter tuning process employs the sparrow search algorithm (SSA) model to enhance the BBN model performance. The performance of the BMLSSA-CAD method is examined using UNSWNB51 and UCI SECOM datasets. The experimental validation of the BMLSSA-CAD method highlighted superior accuracy outcomes of 97.84% and 98.93% compared to recent techniques on the IIoT platform.

PMID:39592667 | DOI:10.1038/s41598-024-79632-4

Sci Rep. 2024 Nov 26;14(1):29397. doi: 10.1038/s41598-024-80239-y.

ABSTRACT

Synthetic Aperture Radar (SAR) integrated with deep learning has been widely used in several military and civilian applications, such as border patrolling, to monitor and regulate the movement of people and goods across land, air, and maritime borders. Amongst these, maritime borders confront different threats and challenges. Therefore, SAR-based ship detection becomes essential for naval surveillance in marine traffic management, oil spill detection, illegal fishing, and maritime piracy. However, the model becomes insensitive to small ships due to the wide-scale variance and uneven distribution of ship sizes in SAR images. This increases the difficulties associated with ship recognition, which triggers several false alarms. To effectively address these difficulties, the present work proposes an ensemble model (eYOLO) based on YOLOv4 and YOLOv5. The model utilizes a weighted box fusion technique to fuse the outputs of YOLOv4 and YOLOv5. Also, a generalized intersection over union loss has been adopted in eYOLO which ensures the increased generalization capability of the model with reduced scale sensitivity. The model has been developed end-to-end, and its performance has been validated against other reported results using an open-source SAR-ship dataset. The obtained results authorize the effectiveness of eYOLO in multi-scale ship detection with an F1 score and mAP of 91.49% and 92.00%, respectively. This highlights the efficacy of eYOLO in multi-scale ship detection using SAR imagery.

PMID:39592646 | DOI:10.1038/s41598-024-80239-y

Comput Biol Med. 2024 Nov 25;184:109399. doi: 10.1016/j.compbiomed.2024.109399. Online ahead of print.

ABSTRACT

Physical rehabilitation plays a critical role in enhancing health outcomes globally. However, the shortage of physiotherapists, particularly in developing countries where the ratio is approximately ten physiotherapists per million people, poses a significant challenge to effective rehabilitation services. The existing literature on rehabilitation often falls short in data representation and the employment of diverse modalities, limiting the potential for advanced therapeutic interventions. To address this gap, This study integrates Computer Vision and Human Activity Recognition (HAR) technologies to support home-based rehabilitation. The study mitigates this gap by exploring various modalities and proposing a framework for data representation. We introduce a novel framework that leverages both Continuous Wavelet Transform (CWT) and Mel-Frequency Cepstral Coefficients (MFCC) for skeletal data representation. CWT is particularly valuable for capturing the time-frequency characteristics of dynamic movements involved in rehabilitation exercises, enabling a comprehensive depiction of both temporal and spectral features. This dual capability is crucial for accurately modelling the complex and variable nature of rehabilitation exercises. In our analysis, we evaluate 20 CNN-based models and one Vision Transformer (ViT) model. Additionally, we propose 12 hybrid architectures that combine CNN-based models with ViT in bi-model and tri-model configurations. These models are rigorously tested on the UI-PRMD and KIMORE benchmark datasets using key evaluation metrics, including accuracy, precision, recall, and F1-score, with 5-fold cross-validation. Our evaluation also considers real-time performance, model size, and efficiency on low-power devices, emphasising practical applicability. The proposed fused tri-model architectures outperform both single-architectures and bi-model configurations, demonstrating robust performance across both datasets and making the fused models the preferred choice for rehabilitation tasks. Our proposed hybrid model, DenMobVit, consistently surpasses state-of-the-art methods, achieving accuracy improvements of 2.9% and 1.97% on the UI-PRMD and KIMORE datasets, respectively. These findings highlight the effectiveness of our approach in advancing rehabilitation technologies and bridging the gap in physiotherapy services.

PMID:39591669 | DOI:10.1016/j.compbiomed.2024.109399

Biomed Pharmacother. 2024 Nov 25;181:117688. doi: 10.1016/j.biopha.2024.117688. Online ahead of print.

ABSTRACT

The dual-specificity tyrosine-phosphorylation-regulated kinase 1 A (DYRK1A) presents a promising therapeutic target for neurological diseases. However, current inhibitors lack selectivity, which can lead to unexpected side effects and increase the difficulty of studying DYRK1A. Therefore, identifying selective inhibitors targeting DYRK1A is essential for reducing side effects and facilitating neurological disease research. This study aimed to discover DYRK1A inhibitors through a screening pipeline incorporating a deep neural network (DNN) model. Herein, we report an optimized model with an accuracy of 0.93 on a testing set. The pipeline was then performed to identify potential DYRK1A inhibitors from the National Cancer Institute (NCI) library. Four novel DYRK1A inhibitors were identified, and compounds NSC657702 and NSC31059 were noteworthy for their potent inhibition, with IC50 values of 50.9 and 39.5 nM, respectively. NSC31059 exhibited exceptional selectivity across 70 kinases. The compounds also significantly reduced DYRK1A-induced tau phosphorylation at key sites associated with the pathology of neurodegenerative diseases. Moreover, they promoted tubulin polymerization, suggesting a role in microtubule stabilization. Cytotoxicity assessments further confirmed the neuronal safety of the compounds. Together, the results demonstrated a promising screening pipeline and novel DYRK1A inhibitors as candidates for further optimization and development.

PMID:39591664 | DOI:10.1016/j.biopha.2024.117688

PLoS One. 2024 Nov 26;19(11):e0313308. doi: 10.1371/journal.pone.0313308. eCollection 2024.

ABSTRACT

Artificial Intelligence (AI)-based deep learning methods for predicting protein structures are reshaping knowledge development and scientific discovery. Recent large-scale application of AI models for protein structure prediction has changed perceptions about complicated biological problems and empowered a new generation of structure-based hypothesis testing. It is well-recognized that proteins have a modular organization according to archetypal folds. However, it is yet to be determined if predicted structures are tuned to one conformation of flexible proteins or if they represent average conformations. Further, whether or not the answer is protein fold-dependent. Therefore, in this study, we analyzed 2878 proteins with at least ten distinct experimental structures available, from which we can estimate protein topological rigidity verses heterogeneity from experimental measurements. We found that AlphaFold v2 (AF2) predictions consistently return one specific form to high accuracy, with 99.68% of distinct folds (n = 623 out of 628) having an experimental structure within 2.5Å RMSD from a predicted structure. Yet, 27.70% and 10.82% of folds (174 and 68 out of 628 folds) have at least one experimental structure over 2.5Å and 5Å RMSD, respectively, from their AI-predicted structure. This information is important for how researchers apply and interpret the output of AF2 and similar tools. Additionally, it enabled us to score fold types according to how homogeneous versus heterogeneous their conformations are. Importantly, folds with high heterogeneity are enriched among proteins which regulate vital biological processes including immune cell differentiation, immune activation, and metabolism. This result demonstrates that a large amount of protein fold flexibility has already been experimentally measured, is vital for critical cellular processes, and is currently unaccounted for in structure prediction databases. Therefore, the structure-prediction revolution begets the protein dynamics revolution!

PMID:39591473 | DOI:10.1371/journal.pone.0313308

PLoS One. 2024 Nov 26;19(11):e0313000. doi: 10.1371/journal.pone.0313000. eCollection 2024.

ABSTRACT

Mitochondria are intracellular organelles that act as powerhouses by breaking down nutrition molecules to produce adenosine triphosphate (ATP) as cellular fuel. They have their own genetic material called mitochondrial DNA. Alterations in mitochondrial DNA can result in primary mitochondrial diseases, including neurodegenerative disorders. Early detection of these abnormalities is crucial in slowing disease progression. With recent advances in data acquisition techniques such as focused ion beam scanning electron microscopy, it has become feasible to capture large intracellular organelle volumes at data rates reaching 4Tb/minute, each containing numerous cells. However, manually segmenting large data volumes (gigapixels) can be time-consuming for pathologists. Therefore, there is an urgent need for automated tools that can efficiently segment mitochondria with minimal user intervention. Our article proposes an ensemble of two automatic segmentation pipelines to predict regions of interest specific to mitochondria. This architecture combines the predicted outputs from both pipelines using an ensemble learning-based entropy-weighted fusion technique. The methodology minimizes the impact of individual predictions and enhances the overall segmentation results. The performance of the segmentation task is evaluated using various metrics, ensuring the reliability of our results. We used four publicly available datasets to evaluate our proposed method's effectiveness. Our proposed fusion method has achieved a high score in terms of the mean Jaccard index and dice coefficient for all four datasets. For instance, in the UroCell dataset, our proposed fusion method achieved scores of 0.9644 for the mean Jaccard index and 0.9749 for the Dice coefficient. The mean error rate and pixel accuracy were 0.0062 and 0.9938, respectively. Later, we compared it with state-of-the-art methods like 2D and 3D CNN algorithms. Our ensemble approach shows promising segmentation efficiency with minimal intervention and can potentially aid in the early detection and mitigation of mitochondrial diseases.

PMID:39591424 | DOI:10.1371/journal.pone.0313000

PLOS Digit Health. 2024 Nov 26;3(11):e0000668. doi: 10.1371/journal.pdig.0000668. eCollection 2024 Nov.

ABSTRACT

Locomotive Syndrome (LS) is defined by decreased walking and standing abilities due to musculoskeletal issues. Early diagnosis is vital as LS can be reversed with appropriate intervention. Although diagnosing LS using standardized charts is straightforward, the labor-intensive and time-consuming nature of the process limits its widespread implementation. To address this, we introduced a Deep Learning (DL)-based computer vision model that employs OpenPose for pose estimation and MS-G3D for spatial-temporal graph analysis. This model objectively assesses gait patterns through single-camera video captures, offering a novel and efficient method for LS prediction and analysis. Our model was trained and validated using a dataset of 186 walking videos, plus 65 additional videos for external validation. The model achieved an average sensitivity of 0.86, demonstrating high effectiveness in identifying individuals with LS. The model's positive predictive value was 0.85, affirming its reliable LS detection, and it reached an overall accuracy rate of 0.77. External validation using an independent dataset confirmed strong generalizability with an Area Under the Curve of 0.75. Although the model accurately diagnosed LS cases, it was less precise in identifying non-LS cases. This study pioneers in diagnosing LS using computer vision technology for pose estimation. Our accessible, non-invasive model serves as a tool that can accurately diagnose the labor-intensive LS tests using only visual assessments, streamlining LS detection and expediting treatment initiation. This significantly improves patient outcomes and marks a crucial advancement in digital health, addressing key challenges in management and care of LS.

PMID:39591393 | DOI:10.1371/journal.pdig.0000668

Ann Am Thorac Soc. 2024 Nov 26. doi: 10.1513/AnnalsATS.202401-022OC. Online ahead of print.

ABSTRACT

RATIONALE: Idiopathic pulmonary fibrosis (IPF) is a complex and heterogeneous disease. Given this, we reasoned that differences in genetic profiles may be associated with unique clinical and radiologic features. Computational image analysis, sometimes referred to as radiomics, provides objective, quantitative assessments of radiologic features in subjects with pulmonary fibrosis.

OBJECTIVE: To determine if the genetic risk profile of patients with IPF identifies unique computational imaging phenotypes.

METHODS: Participants with IPF were included in this study if they had genotype data and CT scans of the chest available for computational image analysis. Extent of lung fibrosis and likelihood of a usual interstitial pneumonia (UIP) pattern were scored automatically by using two separate, previously validated deep learning techniques for CT analysis. UIP pattern was also classified visually by radiologists according to established criteria.

MEASUREMENTS AND MAIN RESULTS: Among 334 participants with IPF, MUC5B, FAM13A and ZKSCAN1 were independently associated with the deep learning-based UIP score. None of the common variants were associated with fibrosis extent by computational imaging. We did not find an association between MUC5B, FAM13A or ZKSCAN1 and visually assessed UIP pattern.

CONCLUSIONS: Select genetic variants are associated with computer-based classification of UIP on CT among patients with IPF. Analysis of radiologic features using deep learning may enhance our ability to identify important genotype-phenotype associations in fibrotic lung diseases.

PMID:39591102 | DOI:10.1513/AnnalsATS.202401-022OC

Toxics. 2024 Nov 11;12(11):809. doi: 10.3390/toxics12110809.

ABSTRACT

A microphysiological system (MPS) is an in vitro culture technology that reproduces the physiological microenvironment and functionality of humans and is expected to be applied for drug screening. In this study, we developed an MPS for the structured culture of human iPSC-derived sensory neurons and then predicted drug-induced neurotoxicity by morphological deep learning. Using human iPSC-derived sensory neurons, after the administration of representative anti-cancer drugs, the toxic effects on soma and axons were evaluated by an AI model with neurite images. Significant toxicity was detected in positive drugs and could be classified by different effects on soma or axons, suggesting that the current method provides an effective evaluation of chemotherapy-induced peripheral neuropathy. The results of neurofilament light chain expression changes in the MPS device also agreed with clinical reports. Therefore, the present MPS combined with morphological deep learning is a useful platform for in vitro peripheral neurotoxicity assessment.

PMID:39590989 | DOI:10.3390/toxics12110809

Tomography. 2024 Nov 18;10(11):1814-1831. doi: 10.3390/tomography10110133.

ABSTRACT

Advancements in artificial intelligence (AI) have significantly transformed the field of abdominal radiology, leading to an improvement in diagnostic and disease management capabilities. This narrative review seeks to evaluate the current standing of AI in abdominal imaging, with a focus on recent literature contributions. This work explores the diagnosis and characterization of hepatobiliary, pancreatic, gastric, colonic, and other pathologies. In addition, the role of AI has been observed to help differentiate renal, adrenal, and splenic disorders. Furthermore, workflow optimization strategies and quantitative imaging techniques used for the measurement and characterization of tissue properties, including radiomics and deep learning, are highlighted. An assessment of how these advancements enable more precise diagnosis, tumor description, and body composition evaluation is presented, which ultimately advances the clinical effectiveness and productivity of radiology. Despite the advancements of AI in abdominal imaging, technical, ethical, and legal challenges persist, and these challenges, as well as opportunities for future development, are highlighted.

PMID:39590942 | DOI:10.3390/tomography10110133

J Imaging. 2024 Nov 5;10(11):281. doi: 10.3390/jimaging10110281.

ABSTRACT

Adverse weather (rain, snow, and fog) can negatively impact computer vision tasks by introducing noise in sensor data; therefore, it is essential to recognize weather conditions for building safe and robust autonomous systems in the agricultural and autonomous driving/drone sectors. The performance degradation in computer vision tasks due to adverse weather depends on the type of weather and the intensity, which influences the amount of noise in sensor data. However, existing weather recognition datasets often lack intensity labels, limiting their effectiveness. To address this limitation, we present VARG, a novel video-based weather recognition dataset with weather intensity labels. The dataset comprises a diverse set of short video sequences collected from various social media platforms and videos recorded by the authors, processed into usable clips, and categorized into three major weather categories, rain, fog, and snow, with three intensity classes: absent/no, moderate, and high. The dataset contains 6742 annotated clips from 1079 videos, with the training set containing 5159 clips and the test set containing 1583 clips. Two sets of annotations are provided for training, the first set to train the models as a multi-label weather intensity classifier and the second set to train the models as a multi-class classifier for three weather scenarios. This paper describes the dataset characteristics and presents an evaluation study using several deep learning-based video recognition approaches for weather intensity prediction.

PMID:39590745 | DOI:10.3390/jimaging10110281

J Imaging. 2024 Nov 3;10(11):280. doi: 10.3390/jimaging10110280.

ABSTRACT

Echocardiography is the gold standard for the comprehensive diagnosis of cardiac septal defects (CSDs). Currently, echocardiography diagnosis is primarily based on expert observation, which is laborious and time-consuming. With digitization, deep learning (DL) can be used to improve the efficiency of the diagnosis. This study presents a real-time end-to-end framework tailored for pediatric ultrasound video analysis for CSD decision-making. The framework employs an advanced real-time architecture based on You Only Look Once (Yolo) techniques for CSD decision-making with high accuracy. Leveraging the state of the art with the Yolov8l (large) architecture, the proposed model achieves a robust performance in real-time processes. It can be observed that the experiment yielded a mean average precision (mAP) exceeding 89%, indicating the framework's effectiveness in accurately diagnosing CSDs from ultrasound (US) videos. The Yolov8l model exhibits precise performance in the real-time testing of pediatric patients from Mohammad Hoesin General Hospital in Palembang, Indonesia. Based on the results of the proposed model using 222 US videos, it exhibits 95.86% accuracy, 96.82% sensitivity, and 98.74% specificity. During real-time testing in the hospital, the model exhibits a 97.17% accuracy, 95.80% sensitivity, and 98.15% specificity; only 3 out of the 53 US videos in the real-time process were diagnosed incorrectly. This comprehensive approach holds promise for enhancing clinical decision-making and improving patient outcomes in pediatric cardiology.

PMID:39590744 | DOI:10.3390/jimaging10110280

Geroscience. 2024 Nov 26. doi: 10.1007/s11357-024-01445-0. Online ahead of print.

ABSTRACT

Recently, a deep learning algorithm (DLA) has been developed to predict the chronological age from retinal images. The Retinal Age Gap (RAG), a deviation between predicted age from retinal images (Retinal Age, RA) and chronological age, correlates with mortality and age-related diseases. This study evaluated the reliability and accuracy of RA predictions and analyzed various factors that may influence them. We analyzed two groups of participants: Intravisit and Intervisit, both imaged by color fundus photography. RA was predicted using an established algorithm. The Intervisit group comprised 26 subjects, imaged in two sessions. The Intravisit group had 41 subjects, of whom each eye was photographed twice in one session. The mean absolute test-retest difference in predicted RA was 2.39 years for Intervisit and 2.13 years for Intravisit, with the latter showing higher prediction variability. The chronological age was predicted accurately from fundus photographs. Subsetting image pairs based on differential image quality reduced test-retest discrepancies by up to 50%, but mean image quality was not correlated with retest outcomes. Marked diurnal oscillations in RA predictions were observed, with a significant overestimation in the afternoon compared to the morning in the Intravisit cohort. The order of image acquisition across imaging sessions did not influence RA prediction and subjective age perception did not predict RAG. Inter-eye consistency exceeded 3 years. Our study is the first to explore the reliability of RA predictions. Consistent image quality enhances retest outcomes. The observed diurnal variations in RA predictions highlight the need for standardized imaging protocols, but RAG could soon be a reliable metric in clinical investigations.

PMID:39589693 | DOI:10.1007/s11357-024-01445-0

Anim Cogn. 2024 Nov 26;27(1):80. doi: 10.1007/s10071-024-01918-x.

ABSTRACT

Previous studies suggest that social learning in bumblebees can occur through second-order conditioning, with conspecifics functioning as first-order reinforcers. However, the behavioural mechanisms underlying bumblebees' acquisition of socially learned associations remain largely unexplored. Investigating these mechanisms requires detailed quantification and analysis of the observation process. Here we designed a new 2D paradigm suitable for simple top-down high-speed video recording and analysed bumblebees' observational learning process using a deep-learning-based pose-estimation framework. Two groups of bumblebees observed live conspecifics foraging from either blue or yellow flowers during a single foraging bout, and were subsequently tested for their socially learned colour preferences. Both groups successfully learned the colour indicated by the demonstrators and spent more time facing rewarding flowers-whether occupied by demonstrators or not-compared to non-rewarding flowers. While both groups showed a negative correlation between time spent facing non-rewarding flowers and learning outcomes, the observer bees in the blue group benefited from time spent facing occupied rewarding flowers, whereas the yellow group showed that time facing unoccupied rewarding flowers by the observer bees positively correlated with their learning outcomes. These results suggest that socially influenced colour preferences are shaped by the interplay of different types of observations rather than merely by observing a conspecific at a single colour. Together, these findings provide direct evidence of the dynamical viewing process of observer bees during social observation, opening up new opportunities for exploring the details of more complex social learning in bumblebees and other insects.

PMID:39589587 | DOI:10.1007/s10071-024-01918-x

Med Phys. 2024 Nov 26. doi: 10.1002/mp.17546. Online ahead of print.

ABSTRACT

BACKGROUND: Bi-parametric magnetic resonance imaging (bpMRI) has demonstrated promising results in prostate cancer (PCa) detection. Vision transformers have achieved competitive performance compared to convolutional neural network (CNN) in deep learning, but they need abundant annotated data for training. Self-supervised learning can effectively leverage unlabeled data to extract useful semantic representations without annotation and its associated costs.

PURPOSE: This study proposes a novel self-supervised learning framework and a transformer model to enhance PCa detection using prostate bpMRI.

METHODS AND MATERIALS: We introduce a novel end-to-end Cross-Shaped windows (CSwin) transformer UNet model, CSwin UNet, to detect clinically significant prostate cancer (csPCa) in prostate bpMRI. We also propose a multitask self-supervised learning framework to leverage unlabeled data and improve network generalizability. Using a large prostate bpMRI dataset (PI-CAI) with 1476 patients, we first pretrain CSwin transformer using multitask self-supervised learning to improve data-efficiency and network generalizability. We then finetune using lesion annotations to perform csPCa detection. We also test the network generalization using a separate bpMRI dataset with 158 patients (Prostate158).

RESULTS: Five-fold cross validation shows that self-supervised CSwin UNet achieves 0.888 ± 0.010 aread under receiver operating characterstics curve (AUC) and 0.545 ± 0.060 Average Precision (AP) on PI-CAI dataset, significantly outperforming four comparable models (nnFormer, Swin UNETR, DynUNet, Attention UNet, UNet). On model generalizability, self-supervised CSwin UNet achieves 0.79 AUC and 0.45 AP, still outperforming all other comparable methods and demonstrating good generalization to external data.

CONCLUSIONS: This study proposes CSwin UNet, a new transformer-based model for end-to-end detection of csPCa, enhanced by self-supervised pretraining to enhance network generalizability. We employ an automatic weighted loss (AWL) to unify pretext tasks, improving representation learning. Evaluated on two multi-institutional public datasets, our method surpasses existing methods in detection metrics and demonstrates good generalization to external data.

PMID:39589390 | DOI:10.1002/mp.17546

Med Phys. 2024 Nov 26. doi: 10.1002/mp.17536. Online ahead of print.

ABSTRACT

BACKGROUND: Voxel-based analysis (VBA) for population level radiotherapy (RT) outcomes modeling requires topology preserving inter-patient deformable image registration (DIR) that preserves tumors on moving images while avoiding unrealistic deformations due to tumors occurring on fixed images.

PURPOSE: We developed a tumor-aware recurrent registration (TRACER) deep learning (DL) method and evaluated its suitability for VBA.

METHODS: TRACER consists of encoder layers implemented with stacked 3D convolutional long short term memory network (3D-CLSTM) followed by decoder and spatial transform layers to compute dense deformation vector field (DVF). Multiple CLSTM steps are used to compute a progressive sequence of deformations. Input conditioning was applied by including tumor segmentations with 3D image pairs as input channels. Bidirectional tumor rigidity, image similarity, and deformation smoothness losses were used to optimize the network in an unsupervised manner. TRACER and multiple DL methods were trained with 204 3D computed tomography (CT) image pairs from patients with lung cancers (LC) and evaluated using (a) Dataset I (N = 308 pairs) with DL segmented LCs, (b) Dataset II (N = 765 pairs) with manually delineated LCs, and (c) Dataset III with 42 LC patients treated with RT.

RESULTS: TRACER accurately aligned normal tissues. It best preserved tumors, indicated by the smallest tumor volume difference of 0.24%, 0.40%, and 0.13 % and mean square error in CT intensities of 0.005, 0.005, 0.004, computed between original and resampled moving image tumors, for Datasets I, II, and III, respectively. It resulted in the smallest planned RT tumor dose difference computed between original and resampled moving images of 0.01 and 0.013 Gy when using a female and a male reference.

CONCLUSIONS: TRACER is a suitable method for inter-patient registration involving LC occurring in both fixed and moving images and applicable to voxel-based analysis methods.

PMID:39589333 | DOI:10.1002/mp.17536

J Acoust Soc Am. 2024 Nov 1;156(5):3574-3587. doi: 10.1121/10.0034362.

ABSTRACT

Because a reference signal is often unavailable in real-world scenarios, reference-free speech quality and intelligibility assessment models are important for many speech processing applications. Despite a great number of deep-learning models that have been applied to build non-intrusive speech assessment approaches and achieve promising performance, studies focusing on the hearing impaired (HI) subjects are limited. This paper presents HASA-Net+, a multi-objective non-intrusive hearing-aid speech assessment model, building upon our previous work, HASA-Net. HASA-Net+ improves HASA-Net in several ways: (1) inclusivity for both normal-hearing and HI listeners, (2) integration with pre-trained speech foundation models and fine-tuning techniques, (3) expansion of predictive capabilities to cover speech quality and intelligibility in diverse conditions, including noisy, denoised, reverberant, dereverberated, and vocoded speech, thereby evaluating its robustness, and (4) validation of the generalization capability using an out-of-domain dataset.

PMID:39589329 | DOI:10.1121/10.0034362

Elife. 2024 Nov 26;13:RP101069. doi: 10.7554/eLife.101069.

ABSTRACT

This work proposes µGUIDE: a general Bayesian framework to estimate posterior distributions of tissue microstructure parameters from any given biophysical model or signal representation, with exemplar demonstration in diffusion-weighted magnetic resonance imaging. Harnessing a new deep learning architecture for automatic signal feature selection combined with simulation-based inference and efficient sampling of the posterior distributions, µGUIDE bypasses the high computational and time cost of conventional Bayesian approaches and does not rely on acquisition constraints to define model-specific summary statistics. The obtained posterior distributions allow to highlight degeneracies present in the model definition and quantify the uncertainty and ambiguity of the estimated parameters.

PMID:39589260 | DOI:10.7554/eLife.101069