Data Brief. 2025 Jun 21;61:111815. doi: 10.1016/j.dib.2025.111815. eCollection 2025 Aug.

ABSTRACT

This dataset presents a curated collection of 1896 high-resolution image pairs extracted from four historically significant Vietnamese films set during the Vietnam War era. Each pair consists of an original color frame and its corresponding grayscale version, generated using the ITU-R BT.601 luminance formula. Designed to support research in historical image restoration and colorization, the dataset serves as a benchmark for evaluating AI-driven colorization techniques. Frames were systematically extracted at 3 s intervals from well-preserved archival footage, followed by manual selection to ensure visual diversity and contextual relevance. The dataset is organized into training, validation, and test sets, enabling researchers to train and assess deep learning models for restoring and colorizing historical imagery. In addition to addressing the challenges posed by aged film quality, temporal degradation, and complex visual content, this dataset contributes to digital heritage preservation by making grayscale historical visuals more accessible and engaging for modern audiences. Potential applications include the development of automated colorization systems, domain adaptation research, and AI-powered video restoration from static images.

PMID:40673188 | PMC:PMC12266529 | DOI:10.1016/j.dib.2025.111815

Transl Lung Cancer Res. 2025 Jun 30;14(6):1929-1944. doi: 10.21037/tlcr-24-1039. Epub 2025 Jun 26.

ABSTRACT

BACKGROUND: The development of growth prediction models for multiple pulmonary ground-glass nodules (GGNs) could help predict their growth patterns and facilitate more precise identification of nodules that require close monitoring or early intervention. Previous studies have demonstrated the indolent growth pattern of GGNs and developed growth prediction models; however, these investigations predominantly focused on solitary GGN. This study aimed to investigate the natural history of multiple pulmonary GGNs and develop and validate growth prediction models based on computed tomography (CT) features, radiomics, and deep learning (DL) as well as compare their predictive performances.

METHODS: Patients with two or more persistent GGNs who underwent CT scans between October 2010 and November 2023 and had at least 3 years of follow-up without radiotherapy, chemotherapy, or surgery were retrospectively reviewed. The growth of GGN is defined as an increase in mean diameter by at least 2 mm, an increase in volume by at least 30%, or the emergence or enlargement of a solid component by at least 2 mm. Based on the interval changes during follow-up, the enrolled patients and GGNs were categorized into growth and non-growth groups. The data were randomly divided into a training set and a validation set at a ratio of 7:3. Clinical model, Radiomics model, DL model, Clinical-Radiomics model, and Clinical-DL model were constructed. Model performance was assessed using the area under the receiver operating characteristic curve (AUC).

RESULTS: A total of 732 GGNs [mean diameter (interquartile range, IQR), 5.5 (4.5-6.5) mm] from 231 patients (mean age 54.1±9.9 years; 26.4% male, 73.6% female) were included. Of the 156 (156/231, 67.5%) patients with GGN growth, the fastest-growing GGN had a volume doubling time (VDT) and mass doubling time (MDT) of 2,285 (IQR, 1,369-3,545) and 2,438 (IQR, 1,361-4,140) days, respectively. Among the growing 272 (272/732, 37.2%) GGNs, the median VDT and MDT were 2,934 (IQR, 1,648-4,491) and 2,875 (IQR, 1,619-5,148) days, respectively. Lobulation (P=0.049), vacuole (P=0.009), initial volume (P=0.01), and mass (P=0.01) were risk factors of GGN growth. The sensitivity and specificity of the Clinical model 1, Clinical model 2, Radiomics, DL, Clinical-Radiomics, and Clinical-DL models were 77.2% and 80.0%, 77.2% and 79.3%, 75.9% and 77.8%, 59.5% and 75.6%, 82.3% and 86.7%, 78.5% and 80.7%, respectively. The AUC for Clinical model 1, Clinical model 2, Radiomics, DL, Clinical-Radiomics, and Clinical-DL models were 0.876, 0.869, 0.845, 0.735, 0.908, and 0.887, respectively.

CONCLUSIONS: Multiple pulmonary GGNs exhibit indolent biological behaviour. The Clinical-Radiomics model demonstrated superior accuracy in predicting the growth of multiple GGNs compared to Clinical, Radiomics, DL, Clinical-DL models.

PMID:40673084 | PMC:PMC12261256 | DOI:10.21037/tlcr-24-1039

Front Neurosci. 2025 Jul 2;19:1555657. doi: 10.3389/fnins.2025.1555657. eCollection 2025.

ABSTRACT

OBJECTIVE: Alzheimer's disease (AD) is mainly identified by cognitive function deterioration. Diagnosing AD at early stages poses significant challenges for both researchers and healthcare professionals due to the subtle nature of early brain changes. Currently, electroencephalography (EEG) is widely used in the study of neurodegenerative diseases. However, most existing research relies solely on functional connectivity methods to infer inter-regional brain connectivity, overlooking the importance of spatial connections. Moreover, many existing approaches fail to fully integrate multi-frequency EEG features, limiting the comprehensive understanding of dynamic brain activity across different frequency bands. This study aims to address these limitations by developing a novel graph-based deep learning model that fully utilizes both functional and structural information from multi-frequency EEG data.

METHODS: This paper introduces a Multi-Frequency EEG data-based Multi-Graph Convolutional Network (MF-MGCN) model for AD diagnosis. This method integrates both functional and structural connectivity to more thoroughly capture the relationships among brain regions. By extracting differential entropy (DE) features from five distinct frequency bands of EEG signals for each segment and using graph convolutional networks (GCNs) to aggregate these features, the model effectively distinguishes between AD and healthy controls (HC).

RESULTS: The outcomes show that the developed model outperforms existing methods, achieving 96.15% accuracy and 98.74% AUC in AD and HC classification.

CONCLUSION: These findings highlight the potential of the MF-MGCN model as a clinical tool for Alzheimer's disease diagnosis. This approach could help clinicians detect Alzheimer's at earlier stages, enabling timely intervention and personalized treatment plans.

PMID:40672873 | PMC:PMC12263931 | DOI:10.3389/fnins.2025.1555657

Front Med (Lausanne). 2025 Jul 2;12:1596726. doi: 10.3389/fmed.2025.1596726. eCollection 2025.

ABSTRACT

Retinal diseases are among the leading causes of blindness worldwide, requiring early detection for effective treatment. Manual interpretation of ophthalmic imaging, such as optical coherence tomography (OCT), is traditionally time-consuming, prone to inconsistencies, and requires specialized expertise in ophthalmology. This study introduces OculusNet, an efficient and explainable deep learning (DL) approach for detecting retinal diseases using OCT images. The proposed method is specifically tailored for complex medical image patterns in OCTs to identify retinal disorders, such as choroidal neovascularization (CNV), diabetic macular edema (DME), and age-related macular degeneration characterized by drusen. The model benefits from Saliency Map visualization, an Explainable AI (XAI) technique, to interpret and explain how it reaches conclusions when identifying retinal disorders. Furthermore, the proposed model is deployed on a web page, allowing users to upload retinal OCT images and receive instant detection results. This deployment demonstrates significant potential for integration into ophthalmic departments, enhancing diagnostic accuracy and efficiency. In addition, to ensure an equitable comparison, a transfer learning approach has been applied to four pre-trained models: VGG19, MobileNetV2, VGG16, and DenseNet-121. Extensive evaluation reveals that the proposed OculusNet model achieves a test accuracy of 95.48% and a validation accuracy of 98.59%, outperforming all other models in comparison. Moreover, to assess the proposed model's reliability and generalizability, the Matthews Correlation Coefficient and Cohen's Kappa Coefficient have been computed, validating that the model can be applied in practical clinical settings to unseen data.

PMID:40672824 | PMC:PMC12263695 | DOI:10.3389/fmed.2025.1596726

Plant Environ Interact. 2025 Jul 15;6(4):e70072. doi: 10.1002/pei3.70072. eCollection 2025 Aug.

ABSTRACT

Littoral soils along Ghana's coastal zones, hosting diverse halophytes with multiple potential applications, contain significant salt content due to seawater influence. This study identified and explored the nutritional, ecological, and medicinal significance of these halophytes, focusing on their salt tolerance and desalination abilities. Deep learning image recognition was employed to identify plant species, followed by a greenhouse experiment on five selected halophytes (Ipomoea aquatica, Lactuca taraxacifolia, Paspalum vaginatum, Sesuvium portulacastrum, and Talinum triangulare) to assess their response to varying salt concentrations (0, 25, and 50 dS/m) and soil types (sea sand and arable soil). High salt concentrations (50 dS/m) generally reduced plant growth rates and biomass accumulation while increasing soil electrical conductivity (EC), total dissolved solids (TDS), and pH. Arable soil improved halophyte Relative Growth Rate (RGR) and performance index (PI) by 5% and 52%, respectively, compared to sea sand. Sesuvium portulacastrum exhibited enhanced PI at elevated salinity and demonstrated superior salt ion accumulation in roots and leaves at 50 dS/m. Both P. vaginatum and S. portulacastrum maintained the highest shoot and root dry weights under increased salinity, whereas S. portulacastrum significantly reduced soil EC, pH, Na, and Cl ion contents compared to other species. Sesuvium portulacastrum reduced several soil salinity indicators significantly compared to other species, highlighting its potential for addressing soil and water salinity issues in affected environments. This study shows the potential of Ghana's halophytes in addressing soil salinity-related challenges.

PMID:40672803 | PMC:PMC12264084 | DOI:10.1002/pei3.70072

Front Plant Sci. 2025 Jul 2;16:1604382. doi: 10.3389/fpls.2025.1604382. eCollection 2025.

ABSTRACT

INTRODUCTION: The state monitoring of tobacco leaves during the curing process is crucial for process control and automation of tobacco agricultural production. While most of the existing research on tobacco leaves state recognition focused on the temporal state of the leaves, the morphological state was often neglected. Moreover, the previous research typically used a limited number of non-industrial images for training, creating a significant disparity with the images encountered in actual applications.

METHODS: To investigate the potential of deep learning algorithms in identifying the morphological states of tobacco leaves in real industrial scenarios, a comprehensive and large-scale dataset was developed in this study. This dataset focused on the states of tobacco leaves in actual bulk curing barn in multiple production areas in China, specifically recognizing the degrees of yellowing, browning, and drying. Then, an efficient deep learning method was proposed based on this dataset to enhance the predictive performance.

RESULTS: The prediction accuracy achieved for the yellowing degree, browning degree, and drying degree were 83.0%, 90.5%, and 75.6% respectively. The overall average accuracy, satisfied the requirements of practical application scenarios with a value of 83%.

DISCUSSION: Our proposed framework effectively enables morphological state recognition in industrial curing, supporting parameter optimization and enhanced tobacco quality.

PMID:40672565 | PMC:PMC12263583 | DOI:10.3389/fpls.2025.1604382

JCO Clin Cancer Inform. 2025 Jul;9:e2500051. doi: 10.1200/CCI-25-00051. Epub 2025 Jul 16.

ABSTRACT

PURPOSE: The present study aimed to investigate the role of body composition as an independent image-derived biomarker for clinical outcome prediction in a clinical trial cohort of patients with relapsed or refractory (rel/ref) diffuse large B-cell lymphoma (DLBCL) treated with loncastuximab tesirine.

MATERIALS AND METHODS: The imaging cohort consisted of positron emission tomography/computed tomography scans of 140 patients with rel/ref DLBCL treated with loncastuximab tesirine in the LOTIS-2 (ClinicalTrials.gov identifier: NCT03589469) trial. Body composition analysis was conducted using both manual and deep learning-based segmentation of three primary tissue compartments-skeletal muscle (SM), subcutaneous fat (SF), and visceral fat (VF)-at the L3 level from baseline CT scans. From these segmented compartments, body composition ratio indices, including SM*/VF*, SF*/VF*, and SM*/(VF*+SF*), were derived. Pearson's correlation analysis was used to examine the agreement between manual and automated segmentation. Logistic regression analyses were used to assess the association between the derived indices and treatment response. Cox regression analyses were used to determine the effect of body composition indices on time-to-event outcomes. Body composition indices were considered as continuous and binary variables defined by cut points. The Kaplan-Meier method was used to estimate progression-free survival (PFS) and overall survival (OS).

RESULTS: The manual and automated SM*/VF* indices, as dichotomized, were significant predictors in univariable and multivariable logistic models for failure to achieve complete metabolic response. The manual SM*/VF* index as dichotomized was significantly associated with PFS, but not OS, in univariable and multivariable Cox models.

CONCLUSION: The pretreatment SM*/VF* body composition index shows promise as a biomarker for patients with rel/ref DLBCL undergoing treatment with loncastuximab tesirine. The proposed deep learning-based approach for body composition analysis demonstrated comparable performance to the manual process, presenting a more cost-effective alternative to conventional methods.

PMID:40669032 | DOI:10.1200/CCI-25-00051

PLoS One. 2025 Jul 16;20(7):e0324393. doi: 10.1371/journal.pone.0324393. eCollection 2025.

ABSTRACT

Skin cancer (SC) is the most prominent form of cancer in humans, with over 1 million new cases reported worldwide each year. Early identification of SC plays a crucial role in effective treatment. However, protecting patient data privacy is a major concern in medical research. Therefore, this study presents a smart framework for classifying SC leveraging deep learning (DL), federated learning (FL) and explainable AI (XAI). We tested the presented framework on two well-known datasets, ISBI2016 and ISBI2017. The data was first preprocessed by several techniques: resizing, normalization, balancing, and augmentation. Six advanced DL algorithms-VGG16, Xception, DenseNet169, InceptionV3, MobileViT, and InceptionResNetV2-were applied for classification tasks. Among these, the DenseNet169 algorithm obtained the highest accuracy of 83.3% in ISBI2016 and 92.67% in ISBI2017. All models were then tested in an FL platform to maintain data privacy. In the FL platform, the VGG16 algorithm showed the best results, with 92.08% accuracy on ISBI2016 and 94% on ISBI2017. To ensure model interpretability, an XAI-based algorithm named Local Interpretable Model-Agnostic Explanations (LIME) was used to explain the predictions of the proposed framework. We believe the proposed framework offers a dependable tool for SC diagnosis while protecting sensitive medical data.

PMID:40668852 | DOI:10.1371/journal.pone.0324393

PLoS One. 2025 Jul 16;20(7):e0325332. doi: 10.1371/journal.pone.0325332. eCollection 2025.

ABSTRACT

To more accurately capture the expression of the English humanistic landscape in agricultural industrial parks under the emerging agricultural paradigm of fish-vegetable symbiosis, and to address the limitations of unscientific evaluation standards and inadequate adaptability in Chinese-English translation within multimodal contexts, this study proposes an intelligent translation evaluation framework based on the SPEAKING model-comprising Setting, Participants, Ends, Act Sequence, Key, Instrumentalities, Norms, and Genre. The study identifies the core elements essential for articulating the English humanistic landscape of agricultural industrial parks and conducts a comprehensive analysis from the dual perspectives of translation accuracy and adaptability. Fish-vegetable symbiosis, an ecological agricultural system integrating aquaculture and plant cultivation, emphasizes resource recycling and ecological synergy. Internationally referred to as the "aquaponics system," this model has become a pivotal direction in sustainable ecological agriculture due to its efficiency and environmental compatibility. This study investigates multimodal translation tasks across text, image, and speech data. It addresses two primary challenges: (1) the absence of robust theoretical grounding in existing translation evaluation systems, which leads to partial and insufficiently contextualized assessments in agricultural industrial park translations; and (2) difficulties in maintaining consistency and readability across multimodal translation tasks, particularly in speech and visual modalities. The proposed optimization model integrates linguistic theory with deep learning techniques, providing a detailed analysis of contextual translation elements. Comparative evaluations are conducted against five prominent translation models: Multilingual T5 (mT5), Multilingual Bidirectional and Auto-Regressive Transformers (mBART), Delta Language Model (DeltaLM), Many-to-Many Multilingual Translation Model-100 (M2M-100), and Marian Machine Translation (MarianMT). Experimental results indicate that the proposed model outperforms existing benchmarks across multiple evaluation metrics. For translation accuracy, the Setting score for text data reaches 96.72, exceeding mT5's 92.35; the Instrumentalities score for image data is 96.11, outperforming DeltaLM's 93.12; and the Ends score for speech data achieves 94.83, surpassing MarianMT's 91.67. In terms of translation adaptability, the Genre score for text data is 96.41, compared to mT5's 93.21; the Key score for image data is 92.78, slightly higher than mBART's 92.12; and the Norms score for speech data is 91.78, exceeding DeltaLM's 90.23. These findings offer both theoretical insights and practical implications for enhancing multimodal translation evaluation systems and optimizing cross-modal translation tasks. The proposed model significantly contributes to improving the accuracy and adaptability of language expression in the context of agricultural landscapes, advancing research in intelligent translation and natural language processing.

PMID:40668851 | DOI:10.1371/journal.pone.0325332

PLoS One. 2025 Jul 16;20(7):e0326081. doi: 10.1371/journal.pone.0326081. eCollection 2025.

ABSTRACT

The new progress of deep learning and natural language processing technology has strongly promoted the development of English grammar error correction. However, the existing methods mostly rely on large-scale corpus, and often ignore the fine syntactic correlation in paragraphs, which limits the efficiency in complex grammar error correction scenarios. In order to break through this bottleneck, this study proposes an innovative method to effectively use syntactic features to improve the quality and accuracy of paragraph-level grammar correction. Firstly, the sentence vector representation is constructed by BERT, and then the syntactic structure is extracted by dependency parsing. Then carry out difference fusion analysis, measure the syntactic differences of adjacent sentences by cosine similarity, identify the significant differences caused by grammatical errors according to the preset threshold, lock the position and type of errors, and input the original sentence vector into the Seq2Seq model based on Transformer. The model focuses on the wrong area by attention mechanism to generate correction suggestions. The preliminary results show that this method is significantly better than the existing grammar error correction system. In CoLA dataset, the accuracy is 0.88, which is three percentage points higher than that of BERT-GC. The accuracy of LCoLE dataset is 0.86, which is ahead of the baseline model. The accuracy of FCE data set is 0.89, which has obvious advantages. The accuracy is improved by 3% to a higher level. It shows the excellent effect of this method in grammar error recognition and correction, and has far-reaching significance in providing accurate error correction suggestions, helping English learners improve their writing ability and ensuring the quality of English writing. This study not only presents a powerful approach to English grammar error correction, but also highlights the key value of syntactic features in optimizing natural language processing applications.

PMID:40668821 | DOI:10.1371/journal.pone.0326081

PLoS One. 2025 Jul 16;20(7):e0327125. doi: 10.1371/journal.pone.0327125. eCollection 2025.

ABSTRACT

Burnt area (BA) mapping is crucial for assessing wildfire impact, guiding restoration efforts, and improving fire management strategies. Accurate BA data helps estimate carbon emissions, biodiversity loss, and land surface properties post-fire changes. In this study, we designed and evaluated two deep learning-based architectures, a Custom UNET and a novel UNET-Gated Recurrent Unit (GRU), for burnt area classification using PlanetScope data over Bandipur, India. Both models demonstrated high accuracy in classifying burnt and unburnt areas. Performance metrics, including Precision, Recall, F1-Score, Accuracy, Mean Intersection over Union (IoU), and Dice Coefficient, revealed that the UNET-GRU hybrid consistently outperformed the Custom UNET, particularly in Recall and spatial overlap metrics. The Receiver Operating Characteristic (ROC) curve indicated excellent classification performance for both models, with the UNET-GRU achieving a higher AUC (0.98) compared to the Custom UNET (0.96). These findings highlight the UNET-GRU's enhanced capacity to handle finer distinctions and capture spatial and contextual features, making it a robust choice for burnt area classification in the study area. While both models avoided overfitting and maintained generalizability, integrating GRU into the UNET architecture proved particularly effective for precise classification and spatial accuracy. Our results highlight the potential of the novel UNET-GRU for burnt area mapping using very high-resolution data.

PMID:40668808 | DOI:10.1371/journal.pone.0327125

PLoS Comput Biol. 2025 Jul 16;21(7):e1013268. doi: 10.1371/journal.pcbi.1013268. eCollection 2025 Jul.

ABSTRACT

Ubiquitination is critical in biomedical research. Predicting ubiquitination sites based on deep learning model have advanced the study of ubiquitination. However, traditional supervised model limits in the scenarios where labels are scarcity across species. To address this issue, we introduce EUP, an online webserver for ubiquitination prediction and model interpretation for multi-species. EUP is constructed by extracting lysine site-dependent features from pretrained language model ESM2. Then, utilizing conditional variational inference to reduce the ESM2 features to a lower-dimensional latent representation. By constructing downstream models built on this latent feature representation, EUP exhibited superior performance in predicting ubiquitination sites across species, while maintaining low inference latency. Furthermore, key features for predicting ubiquitination sites were identified across animals, plants, and microbes. The identification of shared key features that capture evolutionarily conserved traits enhances the interpretability of the EUP model for ubiquitination prediction. EUP is free and available at (https://eup.aibtit.com/).

PMID:40668800 | DOI:10.1371/journal.pcbi.1013268

IEEE J Biomed Health Inform. 2025 Jul 16;PP. doi: 10.1109/JBHI.2025.3589662. Online ahead of print.

ABSTRACT

Accurate prediction of drug-target interactions (DTIs) is crucial for accelerating drug discovery and reducing experimental costs. However, challenges such as sparse interactions and heterogeneous datasets complicate this prediction. In this study, we hypothesize that leveraging nonnegative matrix completion and integrating heterogeneous similarity information from multiple biological views can improve the accuracy, interpretability, and scalability of DTI prediction. To validate this, we propose two multi-view fused nonnegative matrix completion methods that combine three key components: (1) a nonnegative matrix completion framework that avoids heuristic rank selection and ensures biologically interpretable predictions; (2) a linear multi-view fusion mechanism, where weights over multiple drug and target similarity matrices are jointly learned through linearly constrained quadratic programming; and (3) multi-graph Laplacian regularization to preserve structural properties within each view. The optimization is performed using two efficient proximal linearization-incorporated block coordinate descent algorithms. Extensive experiments on four gold-standard datasets and a larger real-world dataset demonstrate that our models consistently outperform state-of-the-art single-view, multi-view and deep learning-based DTI prediction methods. Furthermore, ablation studies confirm the contribution of each model component, and scalability analysis highlights the computational efficiency of our approach.

PMID:40668724 | DOI:10.1109/JBHI.2025.3589662

IEEE J Biomed Health Inform. 2025 Jul 16;PP. doi: 10.1109/JBHI.2025.3589772. Online ahead of print.

ABSTRACT

Parkinson's Disease (PD) is often characterized by abnormal gait patterns, which can be objectively and quantitatively diagnosed using Vertical Ground Reaction Force (VGRF) signals. Previous studies have demonstrated the effectiveness of deep learning in VGRF signal analysis. However, the inherent graph structure of VGRF signals has not been adequately considered, limiting the representation of dynamic gait characteristics. To address this, we propose a Multi-Scale Adaptive Directed Graph Neural Network (MS-ADGNN) approach to distinguish the gaits between Parkinson's patients and healthy controls. This method models the VGRF signal as a multi-scale directed graph, capturing the distribution relationships within the plantar sensors and the dynamic pressure conduction during walking. MS-ADGNN integrates an Adaptive Directed Graph Network (ADGN) unit and a Multi-Scale Temporal Convolutional Network (MSTCN) unit. ADGN extracts spatial features from three scales of the directed graph, effectively capturing local and global connectivity. MSTCN extracts multi-scale temporal features, capturing short to long-term dependencies. The proposed method outperforms existing methods on three widely used datasets. In cross-dataset experiments, the average improvements in terms of accuracy, F1-score, and geometric mean are 2.46$\%$, 1.25$\%$, and 1.11$\%$ respectively. Meanwhile, in 10-fold cross-validation experiments, the improvements are 0.78$\%$, 0.83$\%$, and 0.81$\%$ respectively.

PMID:40668723 | DOI:10.1109/JBHI.2025.3589772

IEEE Trans Image Process. 2025 Jul 16;PP. doi: 10.1109/TIP.2025.3588040. Online ahead of print.

ABSTRACT

Deep learning-based palmprint recognition methods take performance to the next level. However, most current methods rely on samples with clean labels. Noisy labels are difficult to avoid in practical applications and may affect the reliability of models, which poses a big challenge. In this paper, we propose a novel Multi-stage Noisy Label Selection and Correction (MNLSC) framework to address this issue. Three stages are proposed to improve the robustness of palmprint recognition. Clean simple samples are firstly selected based on self-supervised learning. A Fourier-based module is constructed to select clean hard samples. A pototype-based module is further introduced for selecting noisy labels from the remaining samples and correcting them. Finally, the model is trained by using clean and corrected labels to improve the performance. Experiments are conducted on several constrained and unconstrained palmprint databases. The results demonstrate the superiority of our method over other methods in dealing with different noise rates. Compared with the baseline method, the accuracy can be improved by up to 33.45% when there are 60% noisy labels.

PMID:40668719 | DOI:10.1109/TIP.2025.3588040

IEEE Trans Pattern Anal Mach Intell. 2025 Jul 16;PP. doi: 10.1109/TPAMI.2025.3589606. Online ahead of print.

ABSTRACT

Deep learning methods have demonstrated state-of-the-art performance in image restoration, especially when trained on large-scale paired datasets. However, acquiring paired data in real-world scenarios poses a significant challenge. Unsupervised restoration approaches based on generative adversarial networks (GANs) offer a promising solution without requiring paired datasets. Yet, these GAN-based approaches struggle to surpass the performance of conventional unsupervised GAN-based frameworks without significantly modifying model structures or increasing the computational complexity. To address these issues, we propose a self-collaboration (SC) strategy for existing restoration models. This strategy utilizes information from the previous stage as feedback to guide subsequent stages, achieving significant performance improvement without increasing the framework's inference complexity. The SC strategy comprises a prompt learning (PL) module and a restorer ($Res$). It iteratively replaces the previous less powerful fixed restorer $\overline{Res}$ in the PL module with a more powerful $Res$. The enhanced PL module generates better pseudo-degraded/clean image pairs, leading to a more powerful $Res$ for the next iteration. Our SC can significantly improve the $Res$ 's performance by over 1.5dB without adding extra parameters or computational complexity during inference. Meanwhile, existing self-ensemble (SE) and our SC strategies enhance the performance of pre-trained restorers from different perspectives. As SE increases computational complexity during inference, we propose a re-boosting module to the SC (Reb-SC) to improve the SC strategy further by incorporating SE into SC without increasing inference time. This approach further enhances the restorer's performance by approximately 0.3 dB. Additionally, we present a baseline framework that includes parallel generative adversarial branches with complementary "self-synthesis" and "unpaired-synthesis" constraints, ensuring the effectiveness of the training framework. Extensive experimental results on restoration tasks demonstrate that the proposed model performs favorably against existing state-of-the-art unsupervised restoration methods. Source code and trained models are publicly available at: https://github.com/linxin0/RSCP2GAN.

PMID:40668718 | DOI:10.1109/TPAMI.2025.3589606

AJR Am J Roentgenol. 2025 Jul 16. doi: 10.2214/AJR.24.32399. Online ahead of print.

ABSTRACT

Background: Variability in prostate biparametric MRI (bpMRI) interpretation limits diagnostic reliability for prostate cancer (PCa). Artificial intelligence (AI) has potential to reduce this variability and improve diagnostic accuracy. Objective: The objective of this study was to evaluate impact of a deep learning AI model on lesion- and patient-level clinically significant PCa (csPCa) and PCa detection rates and interreader agreement in bpMRI interpretations. Methods: This retrospective, multireader, multicenter study used a balanced incomplete block design for MRI randomization. Six radiologists of varying experience interpreted bpMRI scans with and without AI assistance in alternating sessions. The reference standard for lesion-level detection for cases was whole-mount pathology after radical prostatectomy; for control patients, negative 12-core systematic biopsies. In all, 180 patients (120 in the case group, 60 in the control group) who underwent mpMRI and prostate biopsy or radical prostatectomy between January 2013 and December 2022 were included. Lesion-level sensitivity, PPV, patient-level AUC for csPCa and PCa detection, and interreader agreement in lesion-level PI-RADS scores and size measurements were assessed. Results: AI assistance improved lesion-level PPV (PI-RADS ≥ 3: 77.2% [95% CI, 71.0-83.1%] vs 67.2% [61.1-72.2%] for csPCa; 80.9% [75.2-85.7%] vs 69.4% [63.4-74.1%] for PCa; both p < .001), reduced lesion-level sensitivity (PIRADS ≥ 3: 44.4% [38.6-50.5%] vs 48.0% [42.0-54.2%] for csPCa, p = .01; 41.7% [37.0-47.4%] vs 44.9% [40.5-50.2%] for PCa, p = .01), and no difference in patient-level AUC (0.822 [95% CI, 0.768-0.866] vs 0.832 [0.787-0.868] for csPCa, p = .61; 0.833 [0.782-0.874] vs 0.835 [0.792-0.871] for PCa, p = .91). AI assistance improved interreader agreement for lesion-level PI-RADS scores (κ = 0.748 [95% CI, 0.701-0.796] vs 0.336 [0.288-0.381], p < .001), lesion size measurements (coverage probability of 0.397 [0.376-0.419] vs 0.367 [0.349-0.383], p < .001), and patient-level PI-RADS scores (κ = 0.704 [0.627-0.767] versus 0.507 [0.421-0.584], p < .001). Conclusion: AI improved lesion-level PPV and interreader agreement with slightly lower lesion-level sensitivity. Clinical Impact: AI may enhance consistency and reduce false-positives in bpMRI interpretations. Further optimization is required to improve sensitivity without compromising specificity.

PMID:40668633 | DOI:10.2214/AJR.24.32399

Brief Bioinform. 2025 Jul 2;26(4):bbaf342. doi: 10.1093/bib/bbaf342.

ABSTRACT

Drug-target interactions (DTIs) play a key role in drug development, and accurate prediction can significantly improve the efficiency of this process. Traditional experimental methods are reliable but time-consuming and laborious. With the rapid development of deep learning, many DTI prediction methods have emerged. However, most of these methods only focus on the intrinsic features of drugs and targets, while ignoring the relational features between them. In addition, existing graph-based DTI prediction methods often face the challenge of over-smoothing in graph neural networks (GNNs), which limits their prediction accuracy. To address these issues, we propose NASNet-DTI (Drug-target Interactions Based on Node Adaptation and Similarity Networks), a new framework designed to overcome these limitations. NASNet-DTI uses graph convolutional network to extract features from drug molecules and targets separately, and constructs heterogeneous networks to represent two types of nodes: drugs and targets. The edges in the network describe their multiple relationships: drug-drug, target-target, and drug-target. In the feature learning stage, NASNet-DTI adopts a node adaptive learning strategy to dynamically determine the optimal aggregation depth for each node. This ensures that each node can learn the most discriminative features, which effectively alleviates the over-smoothing problem and improves prediction accuracy. Experimental results show that NASNet-DTI significantly outperforms existing methods on multiple datasets, demonstrating its effectiveness and potential as a powerful tool to advance drug discovery and development.

PMID:40668556 | DOI:10.1093/bib/bbaf342

Brief Bioinform. 2025 Jul 2;26(4):bbaf341. doi: 10.1093/bib/bbaf341.

ABSTRACT

The 3D organization of chromatin is essential for the functioning of cellular processes, including transcriptional regulation, genome integrity, chromatin accessibility, and higher order nuclear architecture. However, detecting anomalous chromatin interactions in spatiotemporal Hi-C data remains a significant challenge. We present HiC4D-SPOT, an unsupervised deep-learning framework that models chromatin dynamics using a ConvLSTM-based autoencoder to identify structural anomalies. Benchmarking results demonstrate high reconstruction fidelity, with Pearson Correlation Coefficient and Spearman Correlation Coefficient values of 0.9, while accurately detecting deviations linked to temporal inconsistencies, topologically associating domain (TAD) and loop perturbations, and significant chromatin remodeling events. HiC4D-SPOT successfully identifies swapped time points in a time-swap experiment, captures simulated TAD and loop disruptions with high confidence scores and statistical significance of 0.01, and detects HERV-H boundary weakening during cardiomyocyte differentiation, as well as cohesin-mediated loop loss and recovery-aligning with experimentally observed chromatin remodeling events. These findings establish HiC4D-SPOT as an efficient tool for analyzing 3D chromatin dynamics, enabling the detection of biologically significant structural anomalies in spatiotemporal Hi-C data.

PMID:40668555 | DOI:10.1093/bib/bbaf341

Brief Bioinform. 2025 Jul 2;26(4):bbaf329. doi: 10.1093/bib/bbaf329.

ABSTRACT

Precision medicine tailors medical procedures to individual genetic overviews and offers transformative solutions for rare genetic conditions. Machine learning (ML) has enhanced genome-based precision medicine (GBPM) by enabling accurate diagnoses, customized treatments, and risk assessments. ML tools, including deep learning and ensemble methods, process high-dimensional genomic data and reveal discoveries in rare diseases. This review analyzes the ML applications in GBPM, emphasizing its role in disease classification, therapeutic optimization, and biomarker discovery. Key challenges, such as computational complexity, data scarcity, and ethical concerns, are discussed alongside advancements such as hybrid ML models and real-time genomic analysis. Security issues, including data breaches and ethical challenges, are addressed. This review identifies future directions, emphasizing the need for comprehensible ML models, increasing data-sharing frameworks, and global collaborations. By integrating the current research, this study provides a comprehensive perspective on the use of ML for rare genetic disorders, paving the way for transformative advancements in precision medicine.

PMID:40668553 | DOI:10.1093/bib/bbaf329