J Transl Med. 2025 Mar 10;23(1):298. doi: 10.1186/s12967-025-06254-3.

ABSTRACT

BACKGROUND: The presence of tumour-infiltrating lymphocytes (TILs) is a well-established prognostic biomarker across multiple cancer types, with higher TIL counts being associated with lower recurrence rates and improved patient survival. We aimed to examine whether an automated intraepithelial TIL (iTIL) assessment could stratify patients by risk, with the ability to generalise across independent patient cohorts, using routine H&E slides of colorectal cancer (CRC). To our knowledge, no other existing fully automated iTIL system has demonstrated this capability.

METHODS: An automated method employing deep neural networks was developed to enumerate iTILs in H&E slides of CRC. The method was applied to a Stage III discovery cohort (n = 353) to identify an optimal threshold of 17 iTILs per-mm2 tumour for stratifying relapse-free survival. Using this threshold, patients from two independent Stage II-III validation cohorts (n = 1070, n = 885) were classified as "TIL-High" or "TIL-Low".

RESULTS: Significant stratification was observed in terms of overall survival for a combined validation cohort univariate (HR 1.67, 95%CI 1.39-2.00; p < 0.001) and multivariate (HR 1.37, 95%CI 1.13-1.66; p = 0.001) analysis. Our iTIL classifier was an independent prognostic factor within proficient DNA mismatch repair (pMMR) Stage II CRC cases with clinical high-risk features. Of these, those classified as TIL-High had outcomes similar to pMMR clinical low risk cases, and those classified TIL-Low had significantly poorer outcomes (univariate HR 2.38, 95%CI 1.57-3.61; p < 0.001, multivariate HR 2.17, 95%CI 1.42-3.33; p < 0.001).

CONCLUSIONS: Our deep learning method is the first fully automated system to stratify patient outcome by analysing TILs in H&E slides of CRC, that has shown generalisation capabilities across multiple independent cohorts.

PMID:40065354 | DOI:10.1186/s12967-025-06254-3

J Dent. 2025 Mar 8:105668. doi: 10.1016/j.jdent.2025.105668. Online ahead of print.

ABSTRACT

OBJECTIVES: Convolutional Neural Networks (CNNs) have long dominated image analysis in dentistry, reaching remarkable results in a range of different tasks. However, Transformer-based architectures, originally proposed for Natural Language Processing, are also promising for dental image analysis. The present study aimed to compare CNNs with Transformers for different image analysis tasks in dentistry.

METHODS: Two CNNs (U-Net, DeepLabV3+), two Hybrids (SwinUNETR, UNETR) and two Transformer-based architectures (TransDeepLab, SwinUnet) were compared on three dental segmentation tasks on different image modalities. Datasets consisted of (1) 1881 panoramic radiographs used for tooth segmentation, (2) 1625 bitewings used for tooth structure segmentation, and (3) 2689 bitewings for caries lesions segmentation. All models were trained and evaluated using 5-fold cross-validation.

RESULTS: CNNs were found to be significantly superior over Hybrids and Transformer-based architectures for all three tasks. (1) Tooth segmentation showed mean±SD F1-Score of 0.89±0.009 for CNNs, 0.86±0.015 for Hybrids and 0.83±0.22 for Transformer-based architectures. (2) In tooth structure segmentation CNNs also outperformed with 0.85±0.008 compared to Hybrids 0.84±0.005 and Transformers 0.83±0.011. (3) Even more pronounced results were found for caries lesions segmentation; 0.49±0.031 for CNNs, 0.39±0.072 for Hybrids and 0.32±0.039 for Transformer-based architectures.

CONCLUSION: CNNs significantly outperformed Transformer-based architectures and their Hybrids on three segmentation tasks (teeth, tooth structures, caries lesions) on varying dental data modalities (panoramic and bitewing radiographs).

PMID:40064460 | DOI:10.1016/j.jdent.2025.105668

Photodiagnosis Photodyn Ther. 2025 Mar 8:104552. doi: 10.1016/j.pdpdt.2025.104552. Online ahead of print.

ABSTRACT

BACKGROUND: Diabetic retinopathy (DR) is a leading cause of visual impairment and blindness worldwide, necessitating early detection and accurate diagnosis. This study proposes a novel framework integrating Generative Adversarial Networks (GANs) for data augmentation, denoising autoencoders for noise reduction, and transfer learning with EfficientNetB0 to enhance the performance of DR classification models.

METHODS: GANs were employed to generate high-quality synthetic retinal images, effectively addressing class imbalance and enriching the training dataset. Denoising autoencoders further improved image quality by reducing noise and eliminating common artifacts such as speckle noise, motion blur, and illumination inconsistencies, providing clean and consistent inputs for the classification model. EfficientNetB0 was fine-tuned on the augmented and denoised dataset.

RESULTS: The framework achieved exceptional classification metrics, including 99.00% accuracy, recall, and specificity, surpassing state-of-the-art methods. The study employed a custom-curated OCT dataset featuring high-resolution and clinically relevant images, addressing challenges such as limited annotated data and noisy inputs.

CONCLUSIONS: Unlike existing studies, our work uniquely integrates GANs, autoencoders, and EfficientNetB0, demonstrating the robustness, scalability, and clinical potential of the proposed framework. Future directions include integrating interpretability tools to enhance clinical adoption and exploring additional imaging modalities to further improve generalizability. This study highlights the transformative potential of deep learning in addressing critical challenges in diabetic retinopathy diagnosis.

PMID:40064432 | DOI:10.1016/j.pdpdt.2025.104552

Am J Ophthalmol. 2025 Mar 8:S0002-9394(25)00119-9. doi: 10.1016/j.ajo.2025.03.007. Online ahead of print.

ABSTRACT

OBJECTIVE: To identify genetic determinants specific to reticular pseudodrusen (RPD) compared with drusen.

DESIGN: Genome-wide association study (GWAS) SUBJECTS: Participants with RPD, drusen, and controls from the UK Biobank (UKBB), a large, multisite, community-based cohort.

METHODS: A deep learning framework analyzed 169,370 optical coherence tomography (OCT) volumes to identify cases and controls within the UKBB. Five retina specialists validated the cohorts using OCT and color fundus photographs. Several GWAS were undertaken utilizing the quantity and presence of RPD and drusen. Genome-wide significance was defined as p<5e-8.

MAIN OUTCOMES MEASURES: Genetic associations were examined with the number of RPD and drusen within 'pure' cases, where only RPD or drusen were present in either eye. A candidate approach assessed 46 previously known AMD loci. Secondary GWAS were conducted for number of RPD and drusen in mixed cases, and binary case-control analyses for pure RPD and pure drusen.

RESULTS: The study included 1,787 participants: 1,037 controls, 361 pure drusen, 66 pure RPD, and 323 mixed cases. The primary pure RPD GWAS identified four genome-wide significant loci: rs11200630 near ARMS2-HTRA1 (p=1.9e-09), rs79641866 at PARD3B (p=1.3e-08), rs143184903 near ITPR1 (p=8.1e-09), and rs76377757 near SLN (p=4.3e-08). The latter three are uncommon variants (minor allele frequency <5%). A significant association at the CFH locus was also observed using a candidate approach (p=1.8e-04). For pure drusen, two loci reached genome-wide significance: rs10801555 at CFH (p=6.0e-33) and rs61871744 at ARMS2-HTRA1 (p=4.2e-20).

CONCLUSIONS: The study highlights a clear association between the ARMS2-HTRA1 locus and higher RPD load. Although the CFH locus association did not achieve genome-wide significance, a suggestive link was observed. Three novel associations unique to RPD were identified, albeit for uncommon genetic variants. Further studies with larger sample sizes are needed to explore these findings.

PMID:40064387 | DOI:10.1016/j.ajo.2025.03.007

Int J Cardiol. 2025 Mar 8:133140. doi: 10.1016/j.ijcard.2025.133140. Online ahead of print.

ABSTRACT

BACKGROUND: We aimed to develop and validate an Artificial Intelligence (AI) model that leverages CCTA and optical coherence tomography (OCT) images for automated analysis of plaque characteristics and coronary function.

METHODS: A total of 100 patients who underwent invasive coronary angiography, OCT, and CCTA before discharge were included in this study. The data were randomly divided into a training set (80 %) and a test set (20 %). The training set, comprising 21,471 tomography images, was used to train a deep-learning convolutional neural network. Subsequently, the AI model was integrated with flow reserve score calculation software developed by Ruixin Medical.

RESULTS: The results from the test set demonstrated excellent agreement between the AI model and OCT analysis for calcified plaque (McNemar test, p = 0.683), non-calcified plaque (McNemar test, p = 0.752), mixed plaque (McNemar test, p = 1.000), and low-attenuation plaque (McNemar test, p = 1.000). Additionally, there was excellent agreement for deep learning-derived minimum lumen diameter (intraclass correlation coefficient [ICC] 0.91, p < 0.001), mean vessel diameter (ICC 0.88, p < 0.001), and percent diameter stenosis (ICC 0.82, p < 0.001). In diagnosing >50 % coronary stenosis, the diagnostic accuracy of the AI model surpassed that of conventional CCTA (AUC 0.98 vs. 0.76, p = 0.008). When compared with quantitative flow fraction, there was excellent agreement between QFR and AI-derived CT-FFR (ICC 0.745, p < 0.0001).

CONCLUSION: Our AI model effectively provides automated analysis of plaque characteristics from CCTA images, with the analysis results showing strong agreement with OCT findings. Moreover, the CT-FFR automatically analyzed by the AI model exhibits high consistency with QFR derived from coronary angiography.

PMID:40064207 | DOI:10.1016/j.ijcard.2025.133140

Comput Biol Med. 2025 Mar 9;189:109941. doi: 10.1016/j.compbiomed.2025.109941. Online ahead of print.

ABSTRACT

OBJECTIVE: To resolve the underestimation problem and investigate the mechanism of the AI model which employed to predict cardiovascular disease (CVD) risk scores from retinal fundus photos.

METHODS: An ordinal regression Deep Learning (DL) model was proposed to predict 10-year CVD risk scores. The mechanism of the DL model in understanding CVD risk was explored using methods such as transfer learning and saliency maps.

RESULTS: Model development was performed using data from 34,652 participants with good-quality fundus photographs from the UK Biobank and a dataset for external validation collected in Australia comprised of 1376 fundus photos of 401 participants with a desktop retinal camera and a portable retinal camera. The mean [SD] risk-level accuracies across cross-validation folds was 0.772 [0.008], while AUROC for over moderate risk was 0.849 [0.005] and the AUROC for high risk was 0.874 [0.007] on the UK Biobank dataset. The risk-level accuracy for images acquired with the desktop camera data was 0.715, and the accuracy for portable camera data was 0.656 on the external dataset.

CONCLUSIONS: The DL model described in this study has minimized the underestimation problem. Our analysis confirms that the DL model learned CVD risk score prediction primarily from age- and sex-related image representation. Model performance was only slightly degraded when features such as the retinal vessels and colours were removed from the images. Our analysis identified some image features associated with high CVD risk status, such as the peripheral small vessels and the macula areas.

PMID:40064120 | DOI:10.1016/j.compbiomed.2025.109941

J Neural Eng. 2025 Mar 10. doi: 10.1088/1741-2552/adbebe. Online ahead of print.

ABSTRACT

Objective. Electroencephalography (EEG) is a widely used neuroimaging technique known for its cost-effectiveness and user-friendliness. However, the presence of various artifacts leads to a poor signal-to-noise ratio, limiting the precision of analyses and applications. The proposed work focuses on the Electromyography (EMG) artifacts, which are among the most challenging biological artifacts. The currently reported EMG artifact cleaning performance largely depends on the data used for validation, and in the case of machine learning approaches, also on the data used for training. The data are typically gathered either by recruiting subjects to perform specific EMG artifact tasks or by integrating existing datasets. Prevailing approaches, however, tend to rely on intuitive, concept-oriented data collection with minimal justification for the selection of artifacts and their quantities. Given the substantial costs associated with biological data collection and the pressing need for effective data utilization, we propose an optimization procedure for data-oriented data collection design using deep learning-based artifact detection.Approach. We apply a binary classification differentiating between artifact epochs (time intervals containing EMG artifacts) and non-artifact epochs (time intervals containing no EMG artifact) using three different neural architectures. Our aim is to minimize data collection efforts while preserving the cleaning efficiency.Main results. We were able to reduce the number of EMG artifact tasks from twelve to three and decrease repetitions of isometric&#xD;contraction tasks from ten to three or sometimes even just one.Significance. Our work addresses the need for effective data utilization in biological data collection, offering a systematic and dynamic quantitative approach. By providing clear justifications for the choices of artifacts and their quantity, we aim to guide future studies toward more effective and economical data collection in EEG and EMG research.

PMID:40064096 | DOI:10.1088/1741-2552/adbebe

J Am Acad Orthop Surg. 2025 Mar 11. doi: 10.5435/JAAOS-D-24-01057. Online ahead of print.

ABSTRACT

Implantation of metallic instrumentation is the mainstay of a variety of orthopaedic and spine surgeries. Postoperatively, imaging of the soft tissues around these implants is commonly required to assess for persistent, recurrent, and/or new pathology (ie, instrumentation loosening, particle disease, infection, neural compression); visualization of these pathologies often requires the superior soft-tissue contrast of magnetic resonance imaging (MRI). As susceptibility artifacts from ferromagnetic implants can result in unacceptable image quality, unique MRI approaches are often necessary to provide accurate imaging. In this text, a comprehensive review is provided on common artifacts encountered in orthopaedic MRI, including comparisons of artifacts from different metallic alloys and common nonpropriety/propriety MR metallic artifact reduction methods. The newest metal-artifact suppression imaging technology and future directions (ie, deep learning/artificial intelligence) in this important field will be considered.

PMID:40063737 | DOI:10.5435/JAAOS-D-24-01057

PLoS One. 2025 Mar 10;20(3):e0317306. doi: 10.1371/journal.pone.0317306. eCollection 2025.

ABSTRACT

Underwater vision is essential in numerous applications, such as marine resource surveying, autonomous navigation, objective detection, and target monitoring. However, raw underwater images often suffer from significant color deviations due to light attenuation, presenting challenges for practical use. This systematic literature review examines the latest advancements in color correction methods for underwater image enhancement. The core objectives of the review are to identify and critically analyze existing approaches, highlighting their strengths, limitations, and areas for future research. A comprehensive search across eight scholarly databases resulted in the identification of 67 relevant studies published between 2010 and 2024. These studies introduce 13 distinct methods for enhancing underwater images, which can be categorized into three groups: physical models, non-physical models, and deep learning-based methods. Physical model-based methods aim to reverse the effects of underwater image degradation by simulating the physical processes of light attenuation and scattering. In contrast, non-physical model-based methods focus on manipulating pixel values without modeling these underlying degradation processes. Deep learning-based methods, by leveraging data-driven approaches, aim to learn mappings between degraded and enhanced images through large datasets. However, challenges persist across all categories, including algorithmic limitations, data dependency, computational complexity, and performance variability across diverse underwater environments. This review consolidates the current knowledge, providing a taxonomy of methods while identifying critical research gaps. It emphasizes the need to improve adaptability across diverse underwater conditions and reduce computational complexity for real-time applications. The review findings serve as a guide for future research to overcome these challenges and advance the field of underwater image enhancement.

PMID:40063649 | DOI:10.1371/journal.pone.0317306

PLoS One. 2025 Mar 10;20(3):e0317630. doi: 10.1371/journal.pone.0317630. eCollection 2025.

ABSTRACT

Portable and wearable electrocardiogram (ECG) devices are increasingly utilized in healthcare for monitoring heart rhythms and detecting cardiac arrhythmias or other heart conditions. The integration of ECG signal visualization with AI-based abnormality detection empowers users to independently and confidently assess their physiological signals. In this study, we investigated a novel method for visualizing ECG signals using polar transformations of short-time Fourier transform (STFT) spectrograms and evaluated the performance of deep convolutional neural networks (CNNs) in predicting atrial fibrillation from these polar transformed spectrograms. The ECG data, which are available from the PhysioNet/CinC Challenge 2017, were categorized into four classes: normal sinus rhythm, atrial fibrillation, other rhythms, and noise. Preprocessing steps included ECG signal filtering, STFT-based spectrogram generation, and reverse polar transformation to generate final polar spectrogram images. These images were used as inputs for deep CNN models, where three pre-trained deep CNNs were used for comparisons. The results demonstrated that deep learning-based predictions using polar transformed spectrograms were comparable to existing methods. Furthermore, the polar transformed images offer a compact and intuitive representation of rhythm characteristics in ECG recordings, highlighting their potential for wearable applications.

PMID:40063554 | DOI:10.1371/journal.pone.0317630

IEEE Trans Vis Comput Graph. 2025 Mar 10;PP. doi: 10.1109/TVCG.2025.3549850. Online ahead of print.

ABSTRACT

As VR technology advances, the demand for multitasking within virtual environments escalates. Negotiating multiple tasks within the immersive virtual setting presents cognitive challenges, where users experience difficulty executing multiple concurrent tasks. This phenomenon highlights the importance of cognitive functions like attention and working memory, which are vital for navigating intricate virtual environments effectively. In addition to attention and working memory, assessing the extent of physical and mental strain induced by the virtual environment and the concurrent tasks performed by the participant is key. While previous research has focused on investigating factors influencing attention and working memory in virtual reality, more comprehensive approaches addressing the prediction of physical and mental strain alongside these cognitive aspects remain. This gap inspired our investigation, where we utilized an open dataset - VRWalking, which included eye and head tracking and physiological measures like heart rate(HR) and galvanic skin response(GSR). The VRwalking dataset has timestamped labeled data for physical and mental load, working memory, and attention metrics. In our investigation, we employed straightforward deep learning models to predict these labels, achieving noteworthy performance with 91%, 96%, 93%, and 91% accuracy in predicting physical load, mental load, working memory, and attention, respectively. Additionally, we conducted SHAP (SHapley Additive exPlanations) analysis to identify the most critical features driving these predictions. Our findings contribute to understanding the overall cognitive state of a participant and effective data collection practices for future researchers, as well as provide insights for virtual reality developers. Developers can utilize these predictive approaches to adaptively optimize user experience in real-time and minimize cognitive strain, ultimately enhancing the effectiveness and usability of virtual reality applications.

PMID:40063446 | DOI:10.1109/TVCG.2025.3549850

IEEE Trans Pattern Anal Mach Intell. 2025 Mar 10;PP. doi: 10.1109/TPAMI.2025.3549300. Online ahead of print.

ABSTRACT

Despite the great success achieved, deep learning technologies usually suffer from data scarcity issues in real-world applications, where existing methods mainly explore sample relationships in a vanilla way from the perspectives of either the input or the loss function. In this paper, we propose a batch transformer module, BatchFormerV1, to equip deep neural networks themselves with the abilities to explore sample relationships in a learnable way. Basically, the proposed method enables data collaboration, e.g., head-class samples will also contribute to the learning of tail classes. Considering that exploring instance-level relationships has very limited impacts on dense prediction, we generalize and refer to the proposed module as BatchFormerV2, which further enables exploring sample relationships for pixel-/patch-level dense representations. In addition, to address the train-test inconsistency where a mini-batch of data samples are neither necessary nor desirable during inference, we also devise a two-stream training pipeline, i.e., a shared model is first jointly optimized with and without BatchFormerV2 which is then removed during testing. The proposed module is plug-and-play without requiring any extra inference cost. Lastly, we evaluate the proposed method on over ten popular datasets, including 1) different data scarcity settings such as long-tailed recognition, zero-shot learning, domain generalization, and contrastive learning; and 2) different visual recognition tasks ranging from image classification to object detection and panoptic segmentation. Code is available at https://zhihou7.github.io/BatchFormer.

PMID:40063428 | DOI:10.1109/TPAMI.2025.3549300

J Fluoresc. 2025 Mar 10. doi: 10.1007/s10895-025-04229-7. Online ahead of print.

ABSTRACT

Camellia oil (CAO), known for its high nutritional and commercial value, has raised increasing concerns about adulteration. Developing an accurate and non-destructive method to identify CAO adulterants is crucial for safeguarding public health and well-being. This study simulates potential real-world adulteration cases by designing representative adulteration scenarios, followed by the acquisition and analysis of corresponding excitation-emission matrix fluorescence (EEMF) spectra. Parallel factor analysis (PARAFAC) was employed to characterize and explore the variations of fluorophores in the EEMF spectra of different adulterated scenarioss, which showed a linear correlation between the relative concentration of PARAFAC components and adulteration levels. A deep learning model named ResTransformer, which combines residual modules with Transformer, was proposed for both the qualitative detection of adulteration types and the quantitative detection of adulteration concentrations from local and global perspectives. The global ResTransformer qualitative models achieved accuracies of over 96.92% based on EEMF spectra and PARAFAC, and quantitative models showed determination coefficient of validation ([Formula: see text]) > 0.978, root mean square error of validation ([Formula: see text]) < 3.0643%, and the ratio performance deviation (RPD) > 7.6741. Compared to traditional chemometric models, the ResTransformer model demonstrated superior performance. The integration of EEMF and ResTransformer presents a highly promising strategy for rapid and reliable detection of CAO adulteration.

PMID:40063235 | DOI:10.1007/s10895-025-04229-7

EJNMMI Phys. 2025 Mar 10;12(1):21. doi: 10.1186/s40658-025-00728-5.

ABSTRACT

PURPOSE: Various specialized and general collimators are used for myocardial perfusion imaging (MPI) with single-photon emission computed tomography (SPECT) to assess different types of coronary artery disease (CAD). Alongside the wide variability in imaging characteristics, the apriori "learnt" information of left ventricular (LV) shape can affect the final diagnosis of the imaging protocol. This study evaluates the effect of prior information incorporation into the segmentation process, compared to deep learning (DL) approaches, as well as the differences of 4 collimation techniques on 5 different datasets.

METHODS: This study was implemented on 80 patients database. 40 patients were coming from mixed black-box collimators, 10 each, from multi-pinhole (MPH), low energy high resolution (LEHR), CardioC and CardioD collimators. The testing was evaluated on a new continuous graph-based approach, which automatically segments the left ventricular volume with prior information on the cardiac geometry. The technique is based on the continuous max-flow (CMF) min-cut algorithm, which performance was evaluated in precision, recall, IoU and Dice score metrics.

RESULTS: In the testing it was shown that, the developed method showed a good improvement over deep learning reaching higher scores in most of the evaluation metrics. Further investigating the different collimation techniques, the evaluation of receiver operating characterstic (ROC) curves showed different stabilities on the various collimators. Running Wilcoxon signed-rank test on the outlines of the LVs showed differentiability between the collimation procedures. To further investigate these phenomena the model parameters of the LVs were reconstructed and evaluated by the uniform manifold approximation and projection (UMAP) method, which further proved that collimators can be differentiated based on the projected LV shapes alone.

CONCLUSIONS: The results show that prior information incorporation can enhance the performance of segmentation methods and collimation strategies have a high effect on the projected cardiac geometry.

PMID:40063231 | DOI:10.1186/s40658-025-00728-5

J Imaging Inform Med. 2025 Mar 10. doi: 10.1007/s10278-025-01470-1. Online ahead of print.

ABSTRACT

Brain tumor is categorized as one of the most fatal form of cancer due to its location and difficulty in terms of diagnostics. Medical expert relies on two key approaches which include biopsy and MRI. However, these techniques have several setbacks which include the need of medical experts, inaccuracy, miss-diagnosis as a result of anxiety or workload which may lead to patient morbidity and mortality. This opens a gap for the need of precise diagnosis and staging to guide appropriate clinical decisions. In this study, we proposed the application of deep learning (DL)-based techniques for the classification of MRI vs non-MRI and tumor vs no tumor. In order to accurately discriminate between classes, we acquired brain tumor multimodal image (CT and MRI) datasets, which comprises of 9616 MRI and CT scans in which 8000 are selected for discrimination between MRI and non-MRI and 4000 for the discrimination between tumor and no tumor cases. The acquired images undergo image pre-processing, data split, data augmentation and model training. The images are trained using 4 DL networks which include MobileNetV2, ResNet, Ineptionv3 and VGG16. Performance evaluation of the DL architectures and comparative analysis has shown that pre-trained MobileNetV2 achieved the best result across all metrics with 99.94% accuracy for the discrimination between MRI and non-MRI and 99.00% for the discrimination between tumor and no tumor. Moreover, I-BrainNet which is a DL/IoT-based framework is developed for the real-time classification of brain tumor.

PMID:40063173 | DOI:10.1007/s10278-025-01470-1

J Imaging Inform Med. 2025 Mar 10. doi: 10.1007/s10278-025-01469-8. Online ahead of print.

ABSTRACT

CT image denoising is a crucial task in medical imaging systems, aimed at enhancing the quality of acquired visual signals. The emergence of diffusion models in machine learning has revolutionized the generation of high-quality CT images. However, diffusion-based CT image denoising methods suffer from two key shortcomings. First, they do not incorporate image formation priors from CT imaging, which limits their adaptability to the CT image denoising task. Second, they are trained on CT images with varying structures and textures at the signal phase, which hinders the model generalization capability. To address the first limitation, we propose a novel conditioning module for our diffusion model that leverages image formation priors from the sinogram domain to generate rich features. To tackle the second issue, we introduce a two-phase training mechanism in which the network gradually learns different anatomical textures and structures. Extensive experimental results demonstrate the effectiveness of both approaches in enhancing CT image quality, with improvements of up to 17% in PSNR and 38% in SSIM, highlighting their superiority over state-of-the-art methods.

PMID:40063172 | DOI:10.1007/s10278-025-01469-8

Int J Cardiovasc Imaging. 2025 Mar 10. doi: 10.1007/s10554-025-03369-y. Online ahead of print.

ABSTRACT

Invasive coronary physiology is underused and carries risks/costs. Artificial Intelligence (AI) might enable non-invasive physiology from invasive coronary angiography (CAG), possibly outperforming humans, but has seldom been explored, especially for instantaneous wave-free Ratio (iFR). We aimed to develop binary iFR lesion classification AI models and compare them with human performance. single-center retrospective study of patients undergoing CAG and iFR. A validated encoder-decoder convolutional neural network (CNN) performed segmentation. Manual annotation of target vessel and pressure sensor location on a segmented telediastolic frame followed. Three AI models classified lesions as positive (≤ 0.89) or negative (> 0.89). Model 1 uses preprocessed vessel diameters with a transformer. Models 2/3 are EfficientNet-B5 CNNs using concatenated angiography and segmentation - Model 3 employs class-frequency-weighted Cross-Entropy Loss. Previous findings demonstrated Model 3's superiority for left anterior descending (LAD) and Model 1's for circumflex (Cx)/right coronary artery (RCA) - they were therefore unified into a vessel-based model. Ten-fold patient-level cross-validation enabled full sample training/testing. Three experienced operators performed binary iFR classification using single frames of raw/segmented images. Comparison metrics were accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Across 250 measurements, AI accuracy was 72%, PPV 48%, NPV 90%, sensitivity 77%, and specificity 71%. Human accuracy ranged from 54 to 74%. NPV was high for the Cx/RCA (AI: 96/98%; operators: 94/97%), but AI significantly outperformed humans in the LAD (78% vs. 60-64%). An AI model capable of binary iFR lesions classification mildly outperformed interventional cardiologists, supporting further validation studies.

PMID:40063156 | DOI:10.1007/s10554-025-03369-y

mSystems. 2025 Mar 10:e0125824. doi: 10.1128/msystems.01258-24. Online ahead of print.

ABSTRACT

Selenoproteins are a special group of proteins with major roles in cellular antioxidant defense. They contain the 21st amino acid selenocysteine (Sec) in the active sites, which is encoded by an in-frame UGA codon. Compared to eukaryotes, identification of selenoprotein genes in bacteria remains challenging due to the absence of an effective strategy for distinguishing the Sec-encoding UGA codon from a normal stop signal. In this study, we have developed a deep learning-based algorithm, deep-Sep, for quickly and precisely identifying selenoprotein genes in bacterial genomic sequences. This algorithm uses a Transformer-based neural network architecture to construct an optimal model for detecting Sec-encoding UGA codons and a homology search-based strategy to remove additional false positives. During the training and testing stages, deep-Sep has demonstrated commendable performance, including an F1 score of 0.939 and an area under the receiver operating characteristic curve of 0.987. Furthermore, when applied to 20 bacterial genomes as independent test data sets, deep-Sep exhibited remarkable capability in identifying both known and new selenoprotein genes, which significantly outperforms the existing state-of-the-art method. Our algorithm has proved to be a powerful tool for comprehensively characterizing selenoprotein genes in bacterial genomes, which should not only assist in accurate annotation of selenoprotein genes in genome sequencing projects but also provide new insights for a deeper understanding of the roles of selenium in bacteria.IMPORTANCESelenium is an essential micronutrient present in selenoproteins in the form of Sec, which is a rare amino acid encoded by the opal stop codon UGA. Identification of all selenoproteins is of vital importance for investigating the functions of selenium in nature. Previous strategies for predicting selenoprotein genes mainly relied on the identification of a special cis-acting Sec insertion sequence (SECIS) element within mRNAs. However, due to the complexity and variability of SECIS elements, recognition of all selenoprotein genes in bacteria is still a major challenge in the annotation of bacterial genomes. We have developed a deep learning-based algorithm to predict selenoprotein genes in bacterial genomic sequences, which demonstrates superior performance compared to currently available methods. This algorithm can be utilized in either web-based or local (standalone) modes, serving as a promising tool for identifying the complete set of selenoprotein genes in bacteria.

PMID:40062874 | DOI:10.1128/msystems.01258-24

Invest Ophthalmol Vis Sci. 2025 Mar 3;66(3):22. doi: 10.1167/iovs.66.3.22.

ABSTRACT

PURPOSE: Loss of retinal perfusion is associated with both onset and worsening of diabetic retinopathy (DR). Optical coherence tomography angiography is a noninvasive method for measuring the nonperfusion area (NPA) and has promise as a scalable screening tool. This study compares two optical coherence tomography angiography algorithms for quantifying NPA.

METHODS: Adults with (N = 101) and without (N = 274) DR were recruited from 20 U.S. sites. We collected 3 × 3-mm macular scans using an Optovue RTVue-XR. Rules-based (RB) and deep-learning-based artificial intelligence (AI) algorithms were used to segment the NPA into four anatomical slabs. For comparison, a subset of scans (n = 50) NPA was graded manually.

RESULTS: The AI method outperformed the RB method in intersection over union, recall, and F1 score, but the RB method has better precision relative to manual grading in all anatomical slabs (all P ≤ 0.001). The AI method had a stronger rank correlation with Early Treatment of Diabetic Retinopathy Study DR severity than the RB method in all slabs (all P < 0.001). NPAs graded using the AI method had a greater area under the receiver operating characteristic curve for diagnosing referable DR than the RB method in the superficial vascular complex, intermediate capillary plexus, and combined inner retina (all P ≤ 0.001), but not in the deep capillary plexus (P = 0.92).

CONCLUSIONS: Our results indicate that output from the AI-based method agrees better with manual grading and can better distinguish between clinically relevant DR severity levels than a RB approach using most plexuses.

PMID:40062815 | DOI:10.1167/iovs.66.3.22

Biochemistry. 2025 Mar 10. doi: 10.1021/acs.biochem.4c00465. Online ahead of print.

ABSTRACT

Chemical peptide engineering, enabled by residue insertion, backbone cyclization, and incorporation of an additional disulfide bond, led to a unique cyclic peptide that efficiently inhibits the invasion of red blood cells by malaria parasites. The engineered peptide exhibits a 20-fold enhanced affinity toward its receptor (PfAMA1) compared to the native peptide ligand (PfRON2), as determined by surface plasmon resonance. In-vitro parasite growth inhibition assay revealed augmented potency of the engineered peptide. The structure of the PfAMA1-cyclic peptide complex, predicted by the deep learning-based structure prediction tool ColabFold-AlphaFold2 with protein-cyclic peptide complex offset, provided valuable insights into the observed activity of the peptide analogs. Rational editing of the peptide backbone and side chain described here proved to be an effective strategy for designing peptide-based inhibitors to interfere with disease-related protein-protein interactions.

PMID:40062812 | DOI:10.1021/acs.biochem.4c00465