J Dent. 2025 Jan 30:105601. doi: 10.1016/j.jdent.2025.105601. Online ahead of print.

ABSTRACT

OBJECTIVES: This study aimed to develop a deep learning (DL) model for the predictive esthetic evaluation of single-implant treatments in the esthetic zone.

METHODS: A total of 226 samples, each comprising three intraoral photographs and 12 clinical features, were collected for proof of concept. Labels were determined by a prosthodontic specialist using the pink esthetic score (PES). A DL model was developed to predict PES based on input images and clinical data. The performance was assessed and compared with that of two other models.

RESULTS: The DL model achieved an average mean absolute error (MAE) of 1.3597, average root mean squared error (MSE) of 1.8324, a Pearson correlation of 0.6326, and accuracies of 65.93% and 85.84% for differences between predicted and ground truth values no larger than 1 and 2, respectively. An ablation study demonstrated that incorporating all input features yielded the best performance, with the proposed model outperforming comparison models.

CONCLUSIONS: DL demonstrates potential for providing acceptable preoperative PES predictions for single implant-supported prostheses in the esthetic zone. Ongoing efforts to collect additional samples and clinical features aim to further enhance the model's performance.

CLINICAL SIGNIFICANCE: The DL model supports dentists in predicting esthetic outcomes and making informed treatment decisions before implant placement. It offers a valuable reference for inexperienced and general dentists to identify esthetic risk factors, thereby improving implant treatment outcomes.

PMID:39892738 | DOI:10.1016/j.jdent.2025.105601

Neuroimage. 2025 Jan 30:121064. doi: 10.1016/j.neuroimage.2025.121064. Online ahead of print.

ABSTRACT

In this study, we present a comprehensive pipeline to train and compare a broad spectrum of machine learning and deep learning brain clocks, integrating diverse preprocessing strategies and correction terms. Our analysis also includes established methodologies which have shown success in prior UK Biobank-related studies. For our analysis we used T1-weighted MRI scans and processed de novo all images via FastSurfer, transforming them into a conformed space for deep learning and extracting image-derived phenotypes for our machine learning approaches. We rigorously evaluated these approaches both as robust age predictors for healthy individuals and as potential biomarkers for various neurodegenerative conditions, leveraging data from the UK Biobank, ADNI, and NACC datasets. To this end we designed a statistical framework to assess age prediction performance, the robustness of the prediction across cohort variability (database, machine type and ethnicity) and its potential as a biomarker for neurodegenerative conditions. Results demonstrate that highly accurate brain age models, typically utilising penalised linear machine learning models adjusted with Zhang's methodology, with mean absolute errors under 1 year in external validation, can be achieved while maintaining consistent prediction performance across different age brackets and subgroups (e.g., ethnicity and MRI machine/manufacturer). Additionally, these models show strong potential as biomarkers for neurodegenerative conditions, such as dementia, where brain age prediction achieved an AUROC of up to 0.90 in distinguishing healthy individuals from those with dementia.

PMID:39892529 | DOI:10.1016/j.neuroimage.2025.121064

Magn Reson Imaging. 2025 Jan 30:110342. doi: 10.1016/j.mri.2025.110342. Online ahead of print.

ABSTRACT

Brain tumor growth is associated with angiogenesis, wherein the density of newly developed blood vessels indicates tumor progression and correlates with the tumor grade. Quantitative dynamic contrast enhanced-magnetic resonance imaging (DCE-MRI) has shown potential in brain tumor grading and treatment response assessment. Segmentation of large-blood-vessels is crucial for automatic and accurate tumor grading using quantitative DCE-MRI. Traditional manual and semi-manual rule-based large-blood-vessel segmentation methods are time-intensive and prone to errors. This study proposes a novel deep learning-based technique for automatic large-blood-vessel segmentation using Swin UNETR architectures and comparing it with U-Net and Attention U-Net architectures. The study employed MRI data from 187 brain tumor patients, with training, validation, and testing datasets sourced from two centers, two vendors, and two field-strength magnetic resonance scanners. To test the generalizability of the developed model, testing was also carried out on different brain tumor types, including lymphoma and metastasis. Performance evaluation demonstrated that Swin UNETR outperformed other models in segmenting large-blood-vessel regions (achieving Dice scores of 0.979, and 0.973 on training and validation sets, respectively, with test set performance ranging from 0.835 to 0.982). Moreover, most quantitative parameters showed significant differences (p < 0.05) between with and without large-blood-vessel. After large-blood-vessel removal, using both ground truth and predicted masks, the values of parameters in non-vascular tumoral regions were statistically similar (p > 0.05). The proposed approach has potential applications in improving the accuracy of automatic grading of tumors as well as in treatment planning.

PMID:39892479 | DOI:10.1016/j.mri.2025.110342

Eur J Radiol. 2025 Jan 29;184:111957. doi: 10.1016/j.ejrad.2025.111957. Online ahead of print.

ABSTRACT

OBJECTIVES: To convert 1D spectra into 2D images using the Gramian angular field, to be used as input for convolutional neural network for classification tasks such as glioblastoma versus lymphoma.

MATERIALS AND METHODS: Retrospective study including patients with histologically confirmed glioblastoma and lymphoma between 2009-2020 who underwent preoperative MR spectroscopy, using single voxel spectroscopy acquired with a short echo time (TE 30). We compared: 1) the Fourier-transformed raw spectra, and 2) the fitted spectra generated during post-processing. Both spectra were independently converted into images using the Gramian angular field, and then served as inputs for a pretrained neural network. We compared the classification performance using data from the Fourier-transformed raw spectra and the post-processed fitted spectra.

RESULTS: This feasibility study included 98 patients, of whom 65 were diagnosed with glioblastomas and 33 with lymphomas. For algorithm testing, 20 % of the cases (19 in total) were randomly selected. By applying the Gramian angular field technique to the Fourier-transformed spectra, we achieved an accuracy of 73.7 % and a sensitivity of 92 % in classifying glioblastoma versus lymphoma, slightly higher than the fitted spectra pathway.

CONCLUSION: Spectroscopy data can be effectively transformed into distinct color graphical images using the Gramian angular field technique, enabling their use as input for deep learning algorithms. Accuracy tends to be higher when utilizing data derived from Fourier-transformed spectra compared to fitted spectra. This finding underscores the potential of using MR spectroscopy data in neural network-based classification purposes.

PMID:39892374 | DOI:10.1016/j.ejrad.2025.111957

Med Image Anal. 2025 Jan 23;101:103474. doi: 10.1016/j.media.2025.103474. Online ahead of print.

ABSTRACT

Current deep learning-based solutions for image analysis tasks are commonly incapable of handling problems to which multiple different plausible solutions exist. In response, posterior-based methods such as conditional Diffusion Models and Invertible Neural Networks have emerged; however, their translation is hampered by a lack of research on adequate validation. In other words, the way progress is measured often does not reflect the needs of the driving practical application. Closing this gap in the literature, we present the first systematic framework for the application-driven validation of posterior-based methods in inverse problems. As a methodological novelty, it adopts key principles from the field of object detection validation, which has a long history of addressing the question of how to locate and match multiple object instances in an image. Treating modes as instances enables us to perform mode-centric validation, using well-interpretable metrics from the application perspective. We demonstrate the value of our framework through instantiations for a synthetic toy example and two medical vision use cases: pose estimation in surgery and imaging-based quantification of functional tissue parameters for diagnostics. Our framework offers key advantages over common approaches to posterior validation in all three examples and could thus revolutionize performance assessment in inverse problems.

PMID:39892221 | DOI:10.1016/j.media.2025.103474

MAGMA. 2025 Feb 1. doi: 10.1007/s10334-024-01222-2. Online ahead of print.

ABSTRACT

OBJECT: Spatio-temporal MRI methods offer rapid whole-brain multi-parametric mapping, yet they are often hindered by prolonged reconstruction times or prohibitively burdensome hardware requirements. The aim of this project is to reduce reconstruction time using deep learning.

MATERIALS AND METHODS: This study focuses on accelerating the reconstruction of volumetric multi-axis spiral projection MRF, aiming for whole-brain T1 and T2 mapping, while ensuring a streamlined approach compatible with clinical requirements. To optimize reconstruction time, the traditional method is first revamped with a memory-efficient GPU implementation. Deep Learning Initialized Compressed Sensing (Deli-CS) is then introduced, which initiates iterative reconstruction with a DL-generated seed point, reducing the number of iterations needed for convergence.

RESULTS: The full reconstruction process for volumetric multi-axis spiral projection MRF is completed in just 20 min compared to over 2 h for the previously published implementation. Comparative analysis demonstrates Deli-CS's efficiency in expediting iterative reconstruction while maintaining high-quality results.

DISCUSSION: By offering a rapid warm start to the iterative reconstruction algorithm, this method substantially reduces processing time while preserving reconstruction quality. Its successful implementation paves the way for advanced spatio-temporal MRI techniques, addressing the challenge of extensive reconstruction times and ensuring efficient, high-quality imaging in a streamlined manner.

PMID:39891798 | DOI:10.1007/s10334-024-01222-2

Environ Monit Assess. 2025 Feb 1;197(2):219. doi: 10.1007/s10661-025-13627-0.

ABSTRACT

Water quality prediction is crucial for water resource management, as accurate forecasting can help identify potential issues in advance and provide a scientific basis for sustainable management. To predict key water quality indicators, including chemical oxygen demand (COD), suspended solids (SS), total phosphorus (TP), pH, total nitrogen (TN), and ammonia nitrogen (NH₃-N), we propose a novel model, CA-TCN-BiGRU, which combines channel attention mechanisms with temporal convolutional networks (TCN) and bidirectional gated recurrent units (BiGRU). The model, which uses a multi-input multi-output (MIMO) architecture, is capable of simultaneously predicting multiple water quality indicators. It is trained and tested using data from a wastewater treatment plant in Huizhou, China. This study investigates the impact of data preprocessing and the channel attention mechanism on model performance and compares the predictive capabilities of various deep learning models. The results demonstrate that data preprocessing significantly improves prediction accuracy, while the channel attention mechanism enhances the model's focus on key features. The CA-TCN-BiGRU model outperforms others in predicting multiple water quality indicators, particularly for COD, TP, and SS, where MAE and RMSE decrease by approximately 23% and 26%, respectively, and R2 improves by 5.85%. Moreover, the model shows strong robustness and real-time performance across different wastewater treatment plants, making it suitable for short-term (1-3 days) water quality prediction. The study concludes that the CA-TCN-BiGRU model not only achieves high accuracy but also offers low computational overhead and fast inference speed, making it an ideal solution for real-time water quality monitoring.

PMID:39891761 | DOI:10.1007/s10661-025-13627-0

Eur Radiol. 2025 Feb 1. doi: 10.1007/s00330-025-11399-2. Online ahead of print.

ABSTRACT

OBJECTIVES: To exploit the capability of super-resolution deep learning reconstruction (SR-DLR) to save radiation exposure from coronary CT angiography (CCTA) and assess its impact on image quality, coronary plaque quantification and characterization, and stenosis severity analysis.

MATERIALS AND METHODS: This prospective study included 50 patients who underwent low-dose (LD) and subsequent ultra-low-dose (ULD) CCTA scans. LD CCTA images were reconstructed with hybrid iterative reconstruction (HIR) and ULD CCTA images were reconstructed with HIR and SR-DLR. The objective parameters and subjective scores were compared. Coronary plaques were classified into three components: necrotic, fibrous or calcified content, with absolute volumes (mm3) recorded, and further characterized by percentage of calcified content. The four main coronary arteries were evaluated for the presence of stenosis. Moreover, 48 coronary segments in 9 patients were evaluated for the presence of significant stenosis, with invasive coronary angiography as a reference.

RESULTS: Effective dose decreased by 60% from LD to ULD CCTA scans (2.01 ± 0.84 mSv vs. 0.80 ± 0.34 mSv, p < 0.001). ULD SR-DLR was non-inferior or even superior to LD HIR in terms of image quality and showed excellent agreements with LD HIR on the plaque volumes, characterization, and stenosis analysis (ICCs > 0.8). Moreover, there was no evidence of a difference in detecting significant coronary stenosis between the LD HIR and ULD SR-DLR (AUC: 0.90 vs. 0.89; p = 1.0).

CONCLUSIONS: SR-DLR led to significant radiation dose savings from CCTA while ensuring high image quality and excellent performance in coronary plaque and stenosis analysis.

KEY POINTS: Question How can radiation dose for coronary CT angiography be reduced without compromising image quality or affecting clinical decisions? Finding Super-resolution deep learning reconstruction (SR-DLR) algorithm allows for 60% dose reduction while ensuring high image quality and excellent performance in coronary plaque and stenosis analysis. Clinical relevance Dose optimization via SR-DLR has no detrimental effect on image quality, coronary plaque quantification and characterization, and stenosis severity analysis, which paves the way for its implementation in clinical practice.

PMID:39891682 | DOI:10.1007/s00330-025-11399-2

Eur Radiol. 2025 Feb 1. doi: 10.1007/s00330-025-11367-w. Online ahead of print.

ABSTRACT

OBJECTIVE: To estimate the developmental trends of quantitative parameters obtained from chest computed tomography (CT) and to provide normative values on dimensions of bronchi and arteries, as well as bronchus-artery (BA) ratios from preschool age to young adulthood.

MATERIALS AND METHODS: Two independent radiologists screened a dataset of 1160 chest CT scans, initially reported as normal, from participants aged 0 to 24 years. Using an automated deep learning-based algorithm, we computed the following bronchus and artery parameters: bronchial outer diameter (Bout), bronchial inner diameter (Bin), adjacent pulmonary artery diameter (A), bronchial wall thickness (Bwt), bronchial wall area (BWA), and bronchial outer area (BOA). From these parameters, we computed the following ratios: Bout/A, Bin/A, Bwt/A, Bwt/Bout, and BWA/BOA. Furthermore, mean lung density, total lung volume, and the square root of wall area of bronchi with a 10-mm lumen perimeter (Pi10) were obtained. The effects on CT parameters of age, sex, and iodine contrast were investigated using mixed-effects or regression model analyses.

RESULTS: 375 normal inspiratory chest CT scans (females / males = 156 / 219; mean age [SD] 12.7 [5.0] years) met the inclusion criteria. Bout and Bin progressively increased with age (all p < 0.05), but Bwt, Bout/A, Bin/A, Bwt/A, Bwt/Bout, or BWA/BOA did not. Total lung volume and mean lung density continuously increased with age (both p < 0.001), while Pi10 did not exhibit such a trend. Bout, total lung volume, and mean lung density were the only parameters that differed between males and females, all higher in males than females (all p < 0.03). The presence of iodinated contrast led to greater values for Bwt, Bwt/Bout, and BWA/BOA, but lower values for Bin, Bout/A, Bin/A, and Bwt/A (all p < 0.01).

CONCLUSION: Quantitative CT parameters of both lung parenchyma and bronchi exhibit growth-related changes, but from 6 to 24 years ratios between bronchus and artery dimensions remain constant. Contrast-enhanced CT scans affect the assessment of lung parenchyma and bronchial size. We propose age and technique-dependent normative values for bronchial dimensions and wall thickness.

KEY POINTS: Question What are the developmental trends of quantitative lung CT parameters in patients from childhood into young adulthood? Findings The ratio between bronchus and pulmonary artery dimensions demonstrates consistent values across age groups, indicating synchronized growth between bronchi and paired pulmonary arteries. Clinical relevance Our findings highlight the importance of standardized CT protocol and volume acquisition, and emphasize the need for ongoing collection of normal chest CT scans to refine the proposed reference values.

PMID:39891681 | DOI:10.1007/s00330-025-11367-w

Neuro Oncol. 2025 Feb 1:noaf026. doi: 10.1093/neuonc/noaf026. Online ahead of print.

ABSTRACT

BACKGROUND: Glioblastoma (GBM) presents a significant clinical challenge due to its aggressive nature and extensive heterogeneity. Tumour purity, the proportion of malignant cells within a tumour, is an important covariate for understanding the disease, having direct clinical relevance or obscuring signal of the malignant portion in molecular analyses of bulk samples. However, current methods for estimating tumour purity are non-specific and technically demanding. Therefore, we aimed to build a reliable and accessible purity estimator for GBM.

METHODS: We developed GBMPurity, a deep-learning model specifically designed to estimate the purity of IDH-wildtype primary GBM from bulk RNA-seq data. The model was trained using simulated pseudobulk tumours of known purity from labelled single-cell data acquired from the GBmap resource. The performance of GBMPurity was evaluated and compared to several existing tools using independent datasets.

RESULTS: GBMPurity outperformed existing tools, achieving a mean absolute error of 0.15 and a concordance correlation coefficient of 0.88 on validation datasets. We demonstrate the utility of GBMPurity through inference on bulk RNA-seq samples and observe reduced purity of the Proneural molecular subtype relative to the Classical, attributed to the increased presence of healthy brain cells.

CONCLUSIONS: GBMPurity provides a reliable and accessible tool for estimating tumour purity from bulk RNA-seq data, enhancing the interpretation of bulk RNA-seq data and offering valuable insights into GBM biology. To facilitate the use of this model by the wider research community, GBMPurity is available as a web-based tool at: https://gbmdeconvoluter.leeds.ac.uk/.

PMID:39891579 | DOI:10.1093/neuonc/noaf026

Bioinformatics. 2025 Jan 31:btaf051. doi: 10.1093/bioinformatics/btaf051. Online ahead of print.

ABSTRACT

MOTIVATION: As the biological roles and disease implications of non-coding RNAs continue to emerge, the need to thoroughly characterize previously unexplored non-coding RNAs becomes increasingly urgent. These molecules hold potential as biomarkers and therapeutic targets. However, the vast and complex nature of non-coding RNAs data presents a challenge. We introduce MMnc, an interpretable deep learning approach designed to classify non-coding RNAs into functional groups. MMnc leverages multiple data sources-such as the sequence, secondary structure, and expression-using attention-based multi-modal data integration. This ensures learning of meaningful representations while accounting for missing sources in some samples.

RESULTS: Our findings demonstrate that MMnc achieves high classification accuracy across diverse non-coding RNA classes. The method's modular architecture allows for the consideration of multiple types of modalities, whereas other tools only consider one or two at most. MMnc is resilient to missing data, ensuring that all available information is effectively utilized. Importantly, the generated attention scores offer interpretable insights into the underlying patterns of the different non-coding RNA classes, potentially driving future non-coding RNA research and applications.

AVAILABILITY: Data and source code can be found at EvryRNA.ibisc.univ-evry.fr/EvryRNA/MMnc.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:39891346 | DOI:10.1093/bioinformatics/btaf051

J Transl Med. 2025 Feb 1;23(1):57. doi: 10.1186/s12967-024-05938-6.

ABSTRACT

BACKGROUND: The traditional process of developing new drugs is time-consuming and often unsuccessful, making drug repurposing an appealing alternative due to its speed and safety. Graph neural networks (GCNs) have emerged as a leading approach for predicting drug-disease associations by integrating drug and disease-related networks with advanced deep learning algorithms. However, GCNs generally infer association probabilities only for existing drugs and diseases, requiring network re-establishment and retraining for novel entities. Additionally, these methods often struggle with sparse networks and fail to elucidate the biological mechanisms underlying newly predicted drugs.

METHODS: To address the limitations of traditional methods, we developed HEDDI-Net, a heterogeneous embedding architecture designed to accurately detect drug-disease associations while preserving the interpretability of biological mechanisms. HEDDI-Net integrates graph and shallow learning techniques to extract representative diseases and proteins, respectively. These representative diseases and proteins are used to embed the input features, which are then utilized in a multilayer perceptron for predicting drug-disease associations.

RESULTS: In experiments, HEDDI-Net achieves areas under the receiver operating characteristic curve of over 0.98, outperforming state-of-the-art methods. Rigorous recovery analyses reveal a median recovery rate of 73% for the top 100 diseases, demonstrating its efficacy in identifying novel target diseases for existing drugs, known as drug repurposing. A case study on Alzheimer's disease highlighted the model's practical applicability and interpretability, identifying potential drug candidates like Baclofen, Fluoxetine, Pentoxifylline and Phenytoin. Notably, over 40% of the predicted candidates in the clusters of commonly prescribed clinical drugs Donepezil and Galantamine had been tested in clinical trials, validating the model's predictive accuracy and practical relevance.

CONCLUSIONS: HEDDI-NET represents a significant advancement by allowing direct application to new diseases and drugs without the need for retraining, a limitation of most GCN-based methods. Furthermore, HEDDI-Net provides detailed affinity patterns with representative proteins for predicted candidate drugs, facilitating an understanding of their physiological effects. This capability also supports the design and testing of alternative drugs that are similar to existing medications, enhancing the reliability and interpretability of potential repurposed drugs. The case study on Alzheimer's disease further underscores HEDDI-Net's ability to predict promising drugs and its applicability in drug repurposing.

PMID:39891114 | DOI:10.1186/s12967-024-05938-6

Eur J Med Res. 2025 Jan 31;30(1):61. doi: 10.1186/s40001-025-02339-3.

ABSTRACT

Malignant gliomas, including glioblastoma, are amongst the most aggressive primary brain tumours, characterised by rapid progression and a poor prognosis. Survival analysis is an essential aspect of glioma management and research, as most studies use time-to-event outcomes to assess overall survival (OS) and progression-free survival (PFS) as key measures to evaluate patients. However, predicting survival using traditional methods such as the Kaplan-Meier estimator and the Cox Proportional Hazards (CPH) model has faced many challenges and inaccuracies. Recently, advances in artificial intelligence (AI), including machine learning (ML) and deep learning (DL), have enabled significant improvements in survival prediction for glioma patients by integrating multimodal data such as imaging, clinical parameters and molecular biomarkers. This study highlights the comparative effectiveness of imaging-based, non-imaging and combined AI models. Imaging models excel at identifying tumour-specific features through radiomics, achieving high predictive accuracy. Non-imaging approaches also excel in utilising clinical and genetic data to provide complementary insights, whilst combined methods integrate multiple data modalities and have the greatest potential for accurate survival prediction. Limitations include data heterogeneity, interpretability challenges and computational demands, particularly in resource-limited settings. Solutions such as federated learning, lightweight AI models and explainable AI frameworks are proposed to overcome these barriers. Ultimately, the integration of advanced AI techniques promises to transform glioma management by enabling personalised treatment strategies and improved prognostic accuracy.

PMID:39891313 | DOI:10.1186/s40001-025-02339-3

BMC Biomed Eng. 2025 Feb 1;7(1):2. doi: 10.1186/s42490-025-00088-2.

ABSTRACT

AIM: The aim of this study is to apply a novel hybrid framework incorporating a Vision Transformer (ViT) and bidirectional long short-term memory (Bi-LSTM) model for classifying physical activity intensity (PAI) in adults using gravity-based acceleration. Additionally, it further investigates how PAI and temporal window (TW) impacts the model' s accuracy.

METHOD: This research used the Capture-24 dataset, consisting of raw accelerometer data from 151 participants aged 18 to 91. Gravity-based acceleration was utilised to generate images encoding various PAIs. These images were subsequently analysed using the ViT-BiLSTM model, with results presented in confusion matrices and compared with baseline models. The model's robustness was evaluated through temporal stability testing and examination of accuracy and loss curves.

RESULT: The ViT-BiLSTM model excelled in PAI classification task, achieving an overall accuracy of 98.5% ± 1.48% across five TWs-98.7% for 1s, 98.1% for 5s, 98.2% for 10s, 99% for 15s, and 98.65% for 30s of TW. The model consistently exhibited superior accuracy in predicting sedentary (98.9% ± 1%) compared to light physical activity (98.2% ± 2%) and moderate-to-vigorous physical activity (98.2% ± 3%). ANOVA showed no significant accuracy variation across PAIs (F = 2.18, p = 0.13) and TW (F = 0.52, p = 0.72). Accuracy and loss curves show the model consistently improves its performance across epochs, demonstrating its excellent robustness.

CONCLUSION: This study demonstrates the ViT-BiLSTM model's efficacy in classifying PAI using gravity-based acceleration, with performance remaining consistent across diverse TWs and intensities. However, PAI and TW could result in slight variations in the model's performance. Future research should concern and investigate the impact of gravity-based acceleration on PAI thresholds, which may influence model's robustness and reliability.

PMID:39891283 | DOI:10.1186/s42490-025-00088-2

Reprod Biol Endocrinol. 2025 Jan 31;23(1):16. doi: 10.1186/s12958-025-01351-w.

ABSTRACT

BACKGROUND: Artificial intelligence (AI) models analyzing embryo time-lapse images have been developed to predict the likelihood of pregnancy following in vitro fertilization (IVF). However, limited research exists on methods ensuring AI consistency and reliability in clinical settings during its development and validation process. We present a methodology for developing and validating an AI model across multiple datasets to demonstrate reliable performance in evaluating blastocyst-stage embryos.

METHODS: This multicenter analysis utilizes time-lapse images, pregnancy outcomes, and morphologic annotations from embryos collected at 10 IVF clinics across 9 countries between 2018 and 2022. The four-step methodology for developing and evaluating the AI model include: (I) curating annotated datasets that represent the intended clinical use case; (II) developing and optimizing the AI model; (III) evaluating the AI's performance by assessing its discriminative power and associations with pregnancy probability across variable data; and (IV) ensuring interpretability and explainability by correlating AI scores with relevant morphologic features of embryo quality. Three datasets were used: the training and validation dataset (n = 16,935 embryos), the blind test dataset (n = 1,708 embryos; 3 clinics), and the independent dataset (n = 7,445 embryos; 7 clinics) derived from previously unseen clinic cohorts.

RESULTS: The AI was designed as a deep learning classifier ranking embryos by score according to their likelihood of clinical pregnancy. Higher AI score brackets were associated with increased fetal heartbeat (FH) likelihood across all evaluated datasets, showing a trend of increasing odds ratios (OR). The highest OR was observed in the top G4 bracket (test dataset G4 score ≥ 7.5: OR 3.84; independent dataset G4 score ≥ 7.5: OR 4.01), while the lowest was in the G1 bracket (test dataset G1 score < 4.0: OR 0.40; independent dataset G1 score < 4.0: OR 0.45). AI score brackets G2, G3, and G4 displayed OR values above 1.0 (P < 0.05), indicating linear associations with FH likelihood. Average AI scores were consistently higher for FH-positive than for FH-negative embryos within each age subgroup. Positive correlations were also observed between AI scores and key morphologic parameters used to predict embryo quality.

CONCLUSIONS: Strong AI performance across multiple datasets demonstrates the value of our four-step methodology in developing and validating the AI as a reliable adjunct to embryo evaluation.

PMID:39891250 | DOI:10.1186/s12958-025-01351-w

BMC Med Inform Decis Mak. 2025 Jan 31;25(1):48. doi: 10.1186/s12911-025-02889-w.

ABSTRACT

This paper introduces SkinWiseNet (SWNet), a deep convolutional neural network designed for the detection and automatic classification of potentially malignant skin cancer conditions. SWNet optimizes feature extraction through multiple pathways, emphasizing network width augmentation to enhance efficiency. The proposed model addresses potential biases associated with skin conditions, particularly in individuals with darker skin tones or excessive hair, by incorporating feature fusion to assimilate insights from diverse datasets. Extensive experiments were conducted using publicly accessible datasets to evaluate SWNet's effectiveness.This study utilized four datasets-Mnist-HAM10000, ISIC2019, ISIC2020, and Melanoma Skin Cancer-comprising skin cancer images categorized into benign and malignant classes. Explainable Artificial Intelligence (XAI) techniques, specifically Grad-CAM, were employed to enhance the interpretability of the model's decisions. Comparative analysis was performed with three pre-existing deep learning networks-EfficientNet, MobileNet, and Darknet. The results demonstrate SWNet's superiority, achieving an accuracy of 99.86% and an F1 score of 99.95%, underscoring its efficacy in gradient propagation and feature capture across various levels. This research highlights the significant potential of SWNet in advancing skin cancer detection and classification, providing a robust tool for accurate and early diagnosis. The integration of feature fusion enhances accuracy and mitigates biases associated with hair and skin tones. The outcomes of this study contribute to improved patient outcomes and healthcare practices, showcasing SWNet's exceptional capabilities in skin cancer detection and classification.

PMID:39891245 | DOI:10.1186/s12911-025-02889-w

J Neural Eng. 2025 Jan 31. doi: 10.1088/1741-2552/adb0f2. Online ahead of print.

ABSTRACT

Abstract-Objective. Steady-state visual evoked potential-based brain-computer interfaces (SSVEP-BCIs) have gained significant attention due to their simplicity, high signal to noise ratio (SNR) and high information transfer rates (ITRs). Currently, accurate detection is a critical issue for enhancing the performance of SSVEP-BCI systems.

APPROACH: This study proposed a novel decoding method called Discriminant Compacted Network (Dis-ComNet), which exploited the advantages of both spatial filtering and deep learning. Specifically, this study enhanced SSVEP features using Global template alignment (GTA) and Discriminant Spatial Pattern (DSP), and then designed a Compacted Temporal-Spatio module (CTSM) to extract finer features. The proposed method was evaluated on a self-collected high-frequency dataset, a public Benchmark dataset and a public wearable dataset.

MAIN RESULTS: The results showed that Dis-ComNet significantly outperformed state-of-the-art spatial filtering methods, deep learning methods, and other fusion methods. Remarkably, Dis-ComNet improved the classification accuracy by 3.9%, 3.5%, 3.2%, 13.3%, 17.4%, 37.5%, and 2.5% when comparing with eTRCA, eTRCA-R, TDCA, DNN, EEGnet, Ensemble-DNN, and TRCA-Net respectively in the high-frequency dataset. The achieved results were 4.7%, 4.6%, 23.6%, 52.5%, 31.7%, and 7.0% higher than those of eTRCA, eTRCA-R, DNN, EEGnet, Ensemble-DNN, and TRCA-Net, respectively, and were comparable to those of TDCA in Benchmark dataset.The accuracy of Dis-ComNet in the wearable dataset was 9.5%, 7.1%, 36.1%, 26.3%, 15.7% and 4.7% higher than eTRCA, eTRCA-R, DNN, EEGnet, Ensemble-DNN, and TRCA-Net respectively, and comparable to TDCA. Besides, our model achieved the ITRs up to 126.0 bits/min, 236.4 bits/min and 103.6 bits/min in the high-frequency, Benchmark and the wearable datasets respectively.

SIGNIFICANCE: This study develops an effective model for the detection of SSVEPs, facilitating the development of high accuracy SSVEP-BCI systems.

PMID:39889306 | DOI:10.1088/1741-2552/adb0f2

J Opt Soc Am A Opt Image Sci Vis. 2024 Jul 1;41(7):1426-1434. doi: 10.1364/JOSAA.523092.

ABSTRACT

Coded aperture compressive temporal imaging (CACTI) is a recently developed imaging technique based on the theory of compressed sensing. It uses an optical imaging system to sample a high-speed dynamic scene (a set of consecutive video frames), integrates the sampled data in time according to masks (sensing matrix), and thus obtains compressive measurements. Considerable effort has been devoted to the sampling strategy and the ill-posed inverse process of reconstructing a three-dimensional (3D) high-speed dynamic scene from two-dimensional (2D) compressive measurements. The importance of the reconstruction algorithm and the optimization mask is evident. In this paper, a flexible, efficient, and superior quality Landweber iterative method is proposed for video reconstruction through jointly learning the optimal binary mask strategy, relaxation strategy, and regularization strategy. To solve the sparse representation problem in iteration, multiple denoisers are introduced to obtain more regularization prior information. By combining the mathematical structure of the Landweber iterative reconstruction method with deep learning, the challenging parameter selection procedure is successfully tackled. Extensive experimental results demonstrate the superiority of the proposed method.

PMID:39889132 | DOI:10.1364/JOSAA.523092

J Opt Soc Am A Opt Image Sci Vis. 2024 Nov 1;41(11):2263-2276. doi: 10.1364/JOSAA.533551.

ABSTRACT

The digital image correlation method is a non-contact optical measurement method, which has the advantages of full-field measurement, simple operation, and high measurement accuracy. The traditional DIC method can accurately measure displacement and strain fields, but there are still many limitations. (i) In the measurement of large displacement deformations, the calculation accuracy of the displacement field and strain field needs to be improved due to the unreasonable setting of parameters such as subset size and step size. (ii) It is difficult to avoid under-matching or over-matching when reconstructing smooth displacement or strain fields. (iii) When processing large-scale image data, the computational complexity will be very high, resulting in slow processing speeds. In recent years, deep-learning-based DIC has shown promising capabilities in addressing the aforementioned issues. We propose a new, to the best of our knowledge, DIC method based on deep learning, which is designed for measuring displacement fields of speckle images in complex large deformations. The network combines the multi-head attention Swin-Transformer and the high-efficient channel attention module ECA and adds positional information to the features to enhance feature representation capabilities. To train the model, we constructed a displacement field dataset that conformed to the real situation and contained various types of speckle images and complex deformations. The measurement results indicate that our model achieves consistent displacement prediction accuracy with traditional DIC methods in practical experiments. Moreover, our model outperforms traditional DIC methods in cases of large displacement scenarios.

PMID:39889089 | DOI:10.1364/JOSAA.533551

J Opt Soc Am A Opt Image Sci Vis. 2024 Sep 1;41(9):1786-1793. doi: 10.1364/JOSAA.526142.

ABSTRACT

Disorders affecting the retina pose a considerable risk to human vision, with an array of factors including aging, diabetes, hypertension, obesity, ocular trauma, and tobacco use exacerbating this issue in contemporary times. Optical coherence tomography (OCT) is a rapidly developing imaging modality that is capable of identifying early signs of vascular, ocular, and central nervous system abnormalities. OCT can diagnose retinal diseases through image classification, but quantifying the laceration area requires image segmentation. To overcome this obstacle, we have developed an innovative deep learning framework that can perform both tasks simultaneously. The suggested framework employs a parallel mask-guided convolutional neural network (PM-CNN) for the classification of OCT B-scans and a grade activation map (GAM) output from the PM-CNN to help a V-Net network (GAM V-Net) to segment retinal lacerations. The guiding mask for the PM-CNN is obtained from the auxiliary segmentation job. The effectiveness of the dual framework was evaluated using a combined dataset that encompassed four publicly accessible datasets along with an additional real-time dataset. This compilation included 11 categories of retinal diseases. The four publicly available datasets provided a robust foundation for the validation of the dual framework, while the real-time dataset enabled the framework's performance to be assessed on a broader range of retinal disease categories. The segmentation Dice coefficient was 78.33±0.15%, while the classification accuracy was 99.10±0.10%. The model's ability to effectively segment retinal fluids and identify retinal lacerations on a different dataset was an excellent demonstration of its generalizability.

PMID:39889044 | DOI:10.1364/JOSAA.526142