Sci Rep. 2025 May 22;15(1):17851. doi: 10.1038/s41598-025-00873-y.

ABSTRACT

In the financial world, Credit card fraud is a budding apprehension in the banking sector, necessitating the development of efficient detection methods to minimize financial losses. The usage of credit cards is experiencing a steady increase, thereby leading to a rise in the default rate that banks encounter. Although there has been much research investigating the efficacy of conventional Machine Learning (ML) models, there has been relatively less emphasis on Deep Learning (DL) techniques. In this article, a machine learning-based system to detect fraudulent transactions using a publicly available dataset of credit card transactions. The dataset, highly imbalanced with fraudulent transactions representing less than 0.2% of the total, was processed using techniques like Synthetic Minority Over-sampling Technique (SMOTE) to handle class imbalance. To predict credit card default, this study evaluates the efficacy of a DL (Deep Learning) model and compares it to other ML models, such as Decision Tree (DT) and Adaboost. The objective of this research is to identify the specific DL parameters that contribute to the observed enhancements in the accuracy of credit card default prediction. This research makes use of the UCI ML repository to access the credit card defaulted customer dataset. Subsequently, various techniques are employed to pre-process the unprocessed data and visually present the outcomes through the use of exploratory data analysis (EDA). Furthermore, the algorithms are hyper tuned to evaluate the enhancement in prediction. We used standard evaluation metrics to evaluate all the models. The evaluation indicates that the Adaboost and DT exhibit the highest accuracy rate of 82 ​% in predicting credit card default, surpassing the accuracy of the ANN model, which is 78 ​%. Several classification algorithms, comprising Logistic Regression, Random Forest, and Neural Networks, were evaluated to determine their effectiveness in identifying fraudulent activities. The Random Forest model emerged as the best performing algorithm with an accuracy of 99.5% and a high recall score, indicating its robustness in detecting fraudulent transactions. This system can be deployed in real-time financial systems to enhance fraud prevention mechanisms and ensure secure financial transactions.

PMID:40404766 | DOI:10.1038/s41598-025-00873-y

Sci Data. 2025 May 22;12(1):843. doi: 10.1038/s41597-025-05211-5.

ABSTRACT

In recent years, deep learning techniques have been extensively used for the identification and classification of lithium-ion batteries. However, these models typically require a costly and labor-intensive labeling process, often influenced by commercial or proprietary concerns. In this study, we introduce RecyBat24, a publicly accessible image dataset for the detection and classification of three battery types: Pouch, Prismatic, and Cylindrical. Our dataset is designed to support both academic research and industrial applications, closely replicating real-world scenarios during the acquisition process and employing data augmentation techniques to simulate various external conditions. Additionally, we demonstrate how the RecyBat24's detection-oriented annotations can be used to create a second version of RecyBat24for instance-segmentation tasks. Finally, we demonstrate that recent lightweight machine learning models achieve high accuracy, highlighting their potential for classification and segmentation applications where computational resources are constrained.

PMID:40404746 | DOI:10.1038/s41597-025-05211-5

Sci Rep. 2025 May 22;15(1):17791. doi: 10.1038/s41598-025-02375-3.

ABSTRACT

This article introduces a novel deep-learning based framework, Super-resolution/Denoising network (SDNet), for simultaneous denoising and super-resolution of swept-source optical coherence tomography (SS-OCT) images. The novelty of this work lies in the hybrid integration of data-driven deep-learning with a model-informed noise representation, specifically designed to address the very low signal-to-noise ratio (SNR) and low-resolution challenges in SS-OCT imaging. SDNet introduces a two-step training process, leveraging noise-free OCT references to simulate low-SNR conditions. In the first step, the network learns to enhance noisy images by combining denoising and super-resolution within noise-corrupted reference domain. To refine its performance, the second step incorporates Principle Component Analysis (PCA) as self-supervised denoising strategy, eliminating the need for ground-truth noisy image data. This unique approach enhances SDNet's adaptability and clinical relevance. A key advantage of SDNet is its ability to balance contrast-texture by adjusting the weights of the two training steps, offering clinicians flexibility for specific diagnostic needs. Experimental results across diverse datasets demonstrate that SDNet surpasses traditional model-based and data-driven methods in computational efficiency, noise reduction, and structural fidelity. The framework excels in improving both image quality and diagnostic accuracy. Additionally, SDNet shows promising adaptability for analyzing low-resolution, low-SNR OCT images, such as those from patients with diabetic macular edema (DME). This study establishes SDNet as a robust, efficient, and clinically adaptable solution for OCT image enhancement addressing critical limitations in contemporary imaging workflows.

PMID:40404743 | DOI:10.1038/s41598-025-02375-3

Clin Radiol. 2025 Apr 24;86:106941. doi: 10.1016/j.crad.2025.106941. Online ahead of print.

ABSTRACT

AIM: Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has proven to be highly sensitive in diagnosing breast tumours, due to the kinetic and volumetric features inherent in it. To utilise the kinetics-related and volume-related information, this paper aims to develop and validate a classification for differentiating benign and malignant breast tumours based on DCE-MRI, though fusing deep features and cross-attention-encoded radiomics features using discriminant correlation analysis (DCA).

MATERIALS AND METHODS: Classification experiments were conducted on a dataset comprising 261 individuals who underwent DCE-MRI including those with multiple tumours, resulting in 137 benign and 163 malignant tumours. To improve the strength of correlation between features and reduce features' redundancy, a novel fusion method that fuses deep features and encoded radiomics features based on DCA (eFF-DCA) is proposed. The eFF-DCA includes three components: (1) a feature extraction module to capture kinetic information across phases, (2) a radiomics feature encoding module employing a cross-attention mechanism to enhance inter-phase feature correlation, and (3) a DCA-based fusion module that transforms features to maximise intra-class correlation while minimising inter-class redundancy, facilitating effective classification.

RESULTS: The proposed eFF-DCA method achieved an accuracy of 90.9% and an area under the receiver operating characteristic curve of 0.942, outperforming methods using single-modal features.

CONCLUSION: The proposed eFF-DCA utilises DCE-MRI kinetic-related and volume-related features to improve breast tumour diagnosis accuracy, but non-end-to-end design limits multimodal fusion. Future research should explore unified end-to-end deep learning architectures that enable seamless multimodal feature fusion and joint optimisation of feature extraction and classification.

PMID:40403340 | DOI:10.1016/j.crad.2025.106941

PLoS One. 2025 May 22;20(5):e0323757. doi: 10.1371/journal.pone.0323757. eCollection 2025.

ABSTRACT

Salient Object Detection (SOD) is a fundamental task in computer vision, aiming to identify prominent regions within images. Traditional methods and deep learning-based models often encounter challenges in capturing crucial information in complex scenes, particularly due to inadequate edge feature extraction, which compromises the precise delineation of object contours and boundaries. To address these challenges, we introduce EFCRFNet, a novel multi-scale feature extraction model that incorporates two innovative modules: the Enhanced Conditional Random Field (ECRF) and the Edge Feature Enhancement Module (EFEM). The ECRF module leverages advanced spatial attention mechanisms to enhance multimodal feature fusion, enabling robust detection in complex environments. Concurrently, the EFEM module focuses on refining edge features to strengthen multi-scale feature representation, significantly improving boundary recognition accuracy. Extensive experiments on standard benchmark datasets demonstrate that EFCRFNet achieves notable performance gains across key evaluation metrics, including MAE (0.64%), Fm (1.04%), Em (8.73%), and Sm (7.4%). These results underscore the effectiveness of EFCRFNet in enhancing detection accuracy and optimizing feature fusion, advancing the state of the art in salient object detection.

PMID:40403088 | DOI:10.1371/journal.pone.0323757

Science. 2025 May 22;388(6749):eadr7094. doi: 10.1126/science.adr7094. Epub 2025 May 22.

ABSTRACT

Deep learning has advanced the design of static protein structures, but the controlled conformational changes that are hallmarks of natural signaling proteins have remained inaccessible to de novo design. Here, we describe a general deep learning-guided approach for de novo design of dynamic changes between intradomain geometries of proteins, similar to switch mechanisms prevalent in nature, with atomic-level precision. We solve four structures that validate the designed conformations, demonstrate modulation of the conformational landscape by orthosteric ligands and allosteric mutations, and show that physics-based simulations are in agreement with deep-learning predictions and experimental data. Our approach demonstrates that new modes of motion can now be realized through de novo design and provides a framework for constructing biology-inspired, tunable, and controllable protein signaling behavior de novo.

PMID:40403060 | DOI:10.1126/science.adr7094

PLoS One. 2025 May 22;20(5):e0324621. doi: 10.1371/journal.pone.0324621. eCollection 2025.

ABSTRACT

As the fashion e-commerce markets rapidly develop, tens of thousands of products are registered daily on e-commerce platforms. Individual sellers register products after setting up a product category directly on a fashion e-commerce platform. However, many sellers fail to find a suitable category and mistakenly register their products under incorrect ones. Precise category matching is important for increasing sales through search optimization and accurate product exposure. However, manually correcting registered categories is time-consuming and costly for platform managers. To resolve this problem, this study proposes a methodology for fashion e-commerce product classification based on multi-modal deep learning and transfer learning. Through the proposed methodology, three challenges in classifying fashion e-commerce products are addressed. First, the issue of extremely biased e-commerce data is addressed through under-sampling. Second, multi-modal deep learning enables the model to simultaneously use input data in different formats, which helps mitigate the impact of noisy and low-quality e-commerce data by providing richer information.Finally, the high computational cost and long training times involved in training deep learning models with both image and text data are mitigated by leveraging transfer learning. In this study, three strategies for transfer learning to fine-tune the image and text modules are presented. In addition, five methods for fusing feature vectors extracted from a single modal into one and six strategies for fine-tuning multi-modal models are presented, featuring a total of 14 strategies. The study shows that multi-modal models outperform unimodal models based solely on text or image. It also suggests the optimal conditions for classifying e-commerce products, helping fashion e-commerce practitioners construct models tailored to their respective business environments more efficiently.

PMID:40403022 | DOI:10.1371/journal.pone.0324621

Database (Oxford). 2025 Apr 22;2025:baaf031. doi: 10.1093/database/baaf031.

ABSTRACT

As the importance of precision agriculture grows, scalable and efficient methods for real-time data collection and analysis have become essential. In this study, we developed a system to collect real-time crop images, focusing on physiological disorders in tomatoes. This system systematically collects crop images and related data, with the potential to evolve into a valuable tool for researchers and agricultural practitioners. A total of 58 479 images were produced under stress conditions, including bacterial wilt (BW), Tomato Yellow Leaf Curl Virus (TYLCV), Tomato Spotted Wilt Virus (TSWV), drought, and salinity, across seven tomato varieties. The images include front views at 0 degrees, 120 degrees, 240 degrees, and top views and petiole images. Of these, 43 894 images were suitable for labeling. Based on this, 24 000 images were used for AI model training, and 13 037 images for model testing. By training a deep learning model, we achieved a mean Average Precision (mAP) of 0.46 and a recall rate of 0.60. Additionally, we discussed data augmentation and hyperparameter tuning strategies to improve AI model performance and explored the potential for generalizing the system across various agricultural environments. The database constructed in this study will serve as a crucial resource for the future development of agricultural AI. Database URL: https://crops.phyzen.com/.

PMID:40402767 | DOI:10.1093/database/baaf031

IEEE J Biomed Health Inform. 2025 May 22;PP. doi: 10.1109/JBHI.2025.3558170. Online ahead of print.

ABSTRACT

Non-suicidal self-injury (NSSI) is a serious threat to the physical and mental health of adolescents, significantly increasing the risk of suicide and attracting widespread public concern. Electroencephalography (EEG), as an objective tool for identifying brain disorders, holds great promise. However, extracting meaningful and reliable features from high-dimensional EEG data, especially by integrating spatiotemporal brain dynamics into informative representations, remains a major challenge. In this study, we introduce an advanced semi-supervised adversarial network, NSSI-Net, to effectively model EEG features related to NSSI. NSSI-Net consists of two key modules: a spatial-temporal feature extraction module and a multi-concept discriminator. In the spatial-temporal feature extraction module, an integrated 2D convolutional neural network (2D-CNN) and a bi-directional Gated Recurrent Unit (BiGRU) are used to capture both spatial and temporal dynamics in EEG data. In the multi-concept discriminator, signal, gender, domain, and disease levels are fully explored to extract meaningful EEG features, considering individual, demographic, disease variations across a diverse population. Based on self-collected NSSI data (n=114), the model's effectiveness and reliability are demonstrated, with a 5.44% improvement in performance compared to existing machine learning and deep learning methods. This study advances the understanding and early diagnosis of NSSI in adolescents with depression, enabling timely intervention.

PMID:40402701 | DOI:10.1109/JBHI.2025.3558170

IEEE J Biomed Health Inform. 2025 May 22;PP. doi: 10.1109/JBHI.2025.3572376. Online ahead of print.

ABSTRACT

This study presents a real-time implementation of an accelerated Hurst Contour Projection from Multiscale Multifractal Analysis (HCP-MMA) for deep learning-based ECG arrhythmia classification. Traditional heart rate variability analyses rely on fixed time scales and predefined parameters, limiting their ability to capture intricate scaling patterns and leading to diagnostic inconsistencies. HCP-MMA converts complex multifractal properties into a contour-based representation, enhancing interpretability for automated classification. However, the high computational cost of MMA hinders real-time processing. To address this, a runtime-optimized parallel computing pipeline is introduced, incorporating singular value decomposition (SVD) and vectorized processing, achieving a $730\times$ speedup over the baseline implementation on an Intel-based system. The proposed HCP-MMA framework, integrated with AlexNet, achieved over 98% classification accuracy across three benchmark datasets (PhysioNet, MIT-BIH, CU), with an F1-score of up to 99.3%. Runtime optimizations enabled real-time deployment on Raspberry Pi 5, demonstrating a $\sim 199\times$ speedup over baseline MMA computation on embedded hardware, with an average inference time of 0.0668 seconds per image, a memory footprint of approximately 220 MB, and a model size of $\sim 122$ MB. Statistical validation using ANOVA and Tukey's HSD tests (p $< 0.05$) confirmed the approach's robustness and generalizability. By bridging computational efficiency with real-time adaptability, this method not only advances automated ECG diagnostics but also paves the way for scalable deployment in wearable monitoring, telemedicine, and multifractal analysis of complex physiological time-series.

PMID:40402700 | DOI:10.1109/JBHI.2025.3572376

IEEE J Biomed Health Inform. 2025 May 22;PP. doi: 10.1109/JBHI.2025.3572458. Online ahead of print.

ABSTRACT

Osteoporotic vertebral compression fractures (OVCFs) are prevalent in the elderly population, typically assessed on computed tomography (CT) scans by evaluating vertebral height loss. This assessment helps determine the fracture's impact on spinal stability and the need for surgical intervention. However, the absence of pre-fracture CT scans and standardized vertebral references leads to measurement errors and inter-observer variability, while irregular compression patterns further challenge the precise grading of fracture severity. While deep learning methods have shown promise in aiding OVCFs screening, they often lack interpretability and sufficient sensitivity, limiting their clinical applicability. To address these challenges, we introduce a novel vertebra synthesis-height loss quantification-OVCFs grading framework. Our proposed model, HealthiVert-GAN, utilizes a coarse-to-fine synthesis network designed to generate pseudo-healthy vertebral images that simulate the pre-fracture state of fractured vertebrae. This model integrates three auxiliary modules that leverage the morphology and height information of adjacent healthy vertebrae to ensure anatomical consistency. Additionally, we introduce the Relative Height Loss of Vertebrae (RHLV) as a quantification metric, which divides each vertebra into three sections to measure height loss between pre-fracture and post-fracture states, followed by fracture severity classification using a Support Vector Machine (SVM). Our approach achieves state-of-the-art classification performance on both the Verse2019 dataset and in-house dataset, and it provides cross-sectional distribution maps of vertebral height loss. This practical tool enhances diagnostic accuracy in clinical settings and assisting in surgical decision-making.

PMID:40402696 | DOI:10.1109/JBHI.2025.3572458

Br J Radiol. 2025 Jun 1;98(1170):974-980. doi: 10.1093/bjr/tqaf059.

ABSTRACT

OBJECTIVES: To investigate the impacts of a deep learning-based iterative reconstruction algorithm on image quality and measuring accuracy of bone mineral density (BMD) in low-dose chest CT.

METHODS: Phantom and patient studies were separately conducted in this study. The same low-dose protocol was used for phantoms and patients. All images were reconstructed with filtered back projection, hybrid iterative reconstruction (HIR) (KARL®, level of 3,5,7), and deep learning-based iterative reconstruction (artificial intelligence iterative reconstruction [AIIR], low, medium, and high strength). The noise power spectrum (NPS) and the task-based transfer function (TTF) were evaluated using phantom. The accuracy and the relative error (RE) of BMD were evaluated using a European spine phantom. The subjective evaluation was performed by 2 experienced radiologists. BMD was measured using quantitative CT (QCT). Image noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), BMD values, and subjective scores were compared with Wilcoxon signed-rank test. The Cohen's kappa test was used to evaluate the inter-reader and inter-group agreement.

RESULTS: AIIR reduced noise and improved resolution on phantom images significantly. There were no significant differences among BMD values in all groups of images (all P > 0.05). RE of BMD measured using AIIR images was smaller. In objective evaluation, all strengths of AIIR achieved less image noise and higher SNR and CNR (all P < 0.05). AIIR-H showed the lowest noise and highest SNR and CNR (P < 0.05). The increase in AIIR algorithm strengths did not affect BMD values significantly (all P > 0.05).

CONCLUSION: The deep learning-based iterative reconstruction did not affect the accuracy of BMD measurement in low-dose chest CT while reducing image noise and improving spatial resolution.

ADVANCES IN KNOWLEDGE: The BMD values could be measured accurately in low-dose chest CT with deep learning-based iterative reconstruction while reducing image noise and improving spatial resolution.

PMID:40402596 | DOI:10.1093/bjr/tqaf059

Br J Radiol. 2025 Jun 1;98(1170):956-964. doi: 10.1093/bjr/tqaf055.

ABSTRACT

OBJECTIVES: To determine the value of on-site deep learning-based CT-derived fractional flow reserve (CT-FFR) for therapeutic management and adverse clinical outcomes in patients suspected of coronary artery disease (CAD) compared with coronary CT angiography (CCTA) alone.

METHODS: This single-centre prospective study included consecutive patients suspected of CAD between June 2021 and September 2021 at our hospital. Four hundred and sixty-one patients were randomized into either CT-FFR+CCTA or CCTA-alone group. The first endpoint was the invasive coronary angiography (ICA) efficiency, defined as the ICA with nonobstructive disease (stenosis <50%) and the ratio of revascularization to ICA (REV-to-ICA ratio) within 90 days. The second endpoint was the incidence of major adverse cardiaovascular events (MACE) at 2 years.

RESULTS: A total of 461 patients (267 [57.9%] men; median age, 64 [55-69]) were included. At 90 days, the rate of ICA with nonobstructive disease in the CT-FFR+CCTA group was lower than in the CCTA group (14.7% vs 34.0%, P=.047). The REV-to-ICA ratio in the CT-FFR+CCTA group was significantly higher than in the CCTA group (73.5% vs. 50.9%, P=.036). No significant difference in ICA efficiency was found in intermediate stenosis (25%-69%) between the 2 groups (all P>.05). After a median follow-up of 23 (22-24) months, MACE were observed in 11 patients in the CT-FFR+CCTA group and 24 in the CCTA group (5.9% vs 10.0%, P=.095).

CONCLUSIONS: The on-site deep learning-based CT-FFR improved the efficiency of ICA utilization with a similarly low rate of MACE compared with CCTA alone.

ADVANCES IN KNOWLEDGE: The on-site deep learning-based CT-FFR was superior to CCTA for therapeutic management.

PMID:40402592 | DOI:10.1093/bjr/tqaf055

JMIR Hum Factors. 2025 May 22;12:e72838. doi: 10.2196/72838.

ABSTRACT

Traditional medical education encounters several challenges. The introduction of advanced facial expression recognition technology offers a new approach to address these issues. The aim of the study is to propose a medical education-assisted teaching and student evaluation method based on facial expression recognition technology. This method consists of 4 key steps. In data collection, multiangle high-definition cameras record students' facial expressions to ensure data comprehensiveness and accuracy. Facial expression recognition uses computer vision and deep learning algorithms to identify students' emotional states. The result analysis stage organizes and statistically analyzes the recognized emotional data to provide teachers with students' learning status feedback. In the teaching feedback stage, teaching strategies are adjusted according to the analysis results. Although this method faces challenges such as technical accuracy, device dependency, and privacy protection, it has the potential to improve teaching effectiveness, optimize personalized learning, and promote teacher-student interaction. The application prospects of this method in medical education are broad, and it is expected to significantly enhance teaching quality and students' learning experience.

PMID:40402552 | DOI:10.2196/72838

Transl Vis Sci Technol. 2025 May 1;14(5):18. doi: 10.1167/tvst.14.5.18.

ABSTRACT

PURPOSE: This study aimed to develop and evaluate a deep learning model for grading foveal hypoplasia using retinal fundus images.

METHODS: This retrospective study included patients with foveal developmental disorders, using color fundus images and optical coherence tomography scans taken between January 1, 2001, and August 31, 2021. In total, 605 retinal fundus images were obtained from 303 patients (male, 55.1%; female, 44.9%). After augmentation, the training, validation, and testing data sets comprised 1229, 527, and 179 images, respectively. A deep learning model was developed for binary classification (normal vs. abnormal foveal development) and six-grade classification of foveal hypoplasia. The outcome was compared with those by senior and junior clinicians.

RESULTS: Higher grade of foveal hypoplasia showed worse visual outcomes (P < 0.001). The binary classification achieved a best testing accuracy of 84.36% using the EfficientNet_b1 model, with 84.51% sensitivity and 84.26% specificity. The six-grade classification achieved a best testing accuracy of 78.21% with the model. The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.9441 and an area under the precision-recall curve (AUPRC) of 0.9654 (both P < 0.0001) in the validation set and an AUROC of 0.8777 and an AUPRC of 0.8327 (both P < 0.0001) in the testing set. Compared to junior and senior clinicians, the EfficientNet_b1 model exhibited a superior performance in both binary and six-grade classification (both P < 0.00001).

CONCLUSIONS: The deep learning model in this study proved more efficient and accurate than assessments by junior and senior clinicians for identifying foveal developmental diseases in retinal fundus images. With the aid of the model, we were able to accurately assess patients with foveal developmental disorders.

TRANSLATIONAL RELEVANCE: This study strengthened the importance for a pediatric deep learning system to support clinical evaluation, particularly in cases reliant on retinal fundus images.

PMID:40402544 | DOI:10.1167/tvst.14.5.18

Clin Transl Oncol. 2025 May 22. doi: 10.1007/s12094-025-03948-4. Online ahead of print.

ABSTRACT

Artificial Intelligence (AI) is transforming neuro-oncology by enhancing diagnosis, treatment planning, and prognosis prediction. AI-driven approaches-such as CNNs and deep learning-have improved the detection and classification of brain tumors through advanced imaging techniques and genomic analysis. Explainable AI methods mitigate the "black box" problem, promoting model transparency and clinical trust. Mechanistic models complement AI by integrating biological principles, enabling precise tumor growth predictions and treatment response assessments. AI applications also include the creation of digital twins for personalized therapy optimization, virtual clinical trials, and predictive modeling for estimation of tumor resection and pattern of recurrence. However, challenges such as data bias, ethical concerns, and regulatory compliance persist. The European Artificial Intelligence Act and the Health Data Space Regulation impose strict data protection and transparency requirements. This review explores AI's methodological foundations, clinical applications, and ethical challenges in neuro-oncology, emphasizing the need for interdisciplinary collaboration and regulatory adaptation.

PMID:40402414 | DOI:10.1007/s12094-025-03948-4

Eur Radiol. 2025 May 22. doi: 10.1007/s00330-025-11689-9. Online ahead of print.

ABSTRACT

OBJECTIVES: To investigate whether a content-based image retrieval (CBIR) of similar chest CT images can help usual interstitial pneumonia (UIP) CT pattern classifications among readers with varying levels of experience.

MATERIALS AND METHODS: This retrospective study included patients who underwent high-resolution chest CT between 2013 and 2015 for the initial workup for fibrosing interstitial lung disease. UIP classifications were assigned to CT images by three thoracic radiologists, which served as the ground truth. One hundred patients were selected as queries. The CBIR retrieved the top three similar CT images with UIP classifications using a deep learning algorithm. The diagnostic accuracies and inter-reader agreement of nine readers before and after CBIR were evaluated.

RESULTS: Of 587 patients (mean age, 63 years; 356 men), 100 query cases (26 UIP patterns, 26 probable UIP patterns, 5 indeterminate for UIP, and 43 alternative diagnoses) were selected. After CBIR, the mean accuracy (61.3% to 67.1%; p = 0.011) and inter-reader agreement (Fleiss Kappa, 0.400 to 0.476; p = 0.003) were slightly improved. The accuracies of the radiologist group for all CT patterns except indeterminate for UIP increased after CBIR; however, they did not reach statistical significance. The resident and pulmonologist groups demonstrated mixed results: accuracy decreased for UIP pattern, increased for alternative diagnosis, and varied for others.

CONCLUSION: CBIR slightly improved diagnostic accuracy and inter-reader agreement in UIP pattern classifications. However, its impact varied depending on the readers' level of experience, suggesting that the current CBIR system may be beneficial when used to complement the interpretations of experienced readers.

KEY POINTS: Question CT pattern classification is important for the standardized assessment and management of idiopathic pulmonary fibrosis, but requires radiologic expertise and shows inter-reader variability. Findings CBIR slightly improved diagnostic accuracy and inter-reader agreement for UIP CT pattern classifications overall. Clinical relevance The proposed CBIR system may guide consistent work-up and treatment strategies by enhancing accuracy and inter-reader agreement in UIP CT pattern classifications by experienced readers whose expertise and experience can effectively interact with CBIR results.

PMID:40402291 | DOI:10.1007/s00330-025-11689-9

Eur Radiol. 2025 May 22. doi: 10.1007/s00330-025-11707-w. Online ahead of print.

ABSTRACT

OBJECTIVES: This study aimed to compare the diagnostic performance of CT-derived fractional flow reserve (CT-FFR) using model-based iterative reconstruction (MBIR) and high-resolution deep learning reconstruction (HR-DLR) images to detect functionally significant stenosis with invasive FFR as the reference standard.

MATERIALS AND METHODS: This single-center retrospective study included 79 consecutive patients (mean age, 70 ± 11 [SD] years; 57 male) who underwent coronary CT angiography followed by invasive FFR between February 2022 and March 2024. CT-FFR was calculated using a mesh-free simulation. The cutoff for functionally significant stenosis was defined as FFR ≤ 0.80. CT-FFR was compared with MBIR and HR-DLR using receiver operating characteristic curve analysis.

RESULTS: The mean invasive FFR value was 0.81 ± 0.09, and 46 of 98 vessels (47%) had FFR ≤ 0.80. The mean noise of HR-DLR was lower than that of MBIR (14.4 ± 1.7 vs 23.5 ± 3.1, p < 0.001). The area under the receiver operating characteristic curve for the diagnosis of functionally significant stenosis of HR-DLR (0.88; 95% CI: 0.80, 0.95) was higher than that of MBIR (0.76; 95% CI: 0.67, 0.86; p = 0.003). The diagnostic accuracy of HR-DLR (88%; 86 of 98 vessels; 95% CI: 80, 94) was higher than that of MBIR (70%; 69 of 98 vessels; 95% CI: 60, 79; p < 0.001).

CONCLUSIONS: HR-DLR improves image quality and the diagnostic performance of CT-FFR for the diagnosis of functionally significant stenosis.

KEY POINTS: Question The effect of HR-DLR on the diagnostic performance of CT-FFR has not been investigated. Findings HR-DLR improved the diagnostic performance of CT-FFR over MBIR for the diagnosis of functionally significant stenosis as assessed by invasive FFR. Clinical relevance HR-DLR would further enhance the clinical utility of CT-FFR in diagnosing the functional significance of coronary stenosis.

PMID:40402290 | DOI:10.1007/s00330-025-11707-w

Int J Legal Med. 2025 May 22. doi: 10.1007/s00414-025-03497-z. Online ahead of print.

ABSTRACT

The epiphyses of the hand and wrist serve as crucial indicators for assessing skeletal maturity in adolescents. This study aimed to develop a deep learning (DL) model for bone age (BA) assessment using hand and wrist X-ray images, addressing the challenge of classifying BA in adolescents. The results of this DL-based classification were then compared and analyzed with those obtained from manual assessment. A retrospective analysis was conducted on 688 hand and wrist X-ray images of adolescents aged 11.00-23.99 years from western China, which were randomly divided into training set, validation set and test set. The BA assessment results were initially analyzed and compared using four DL network models: InceptionV3, InceptionV3 + SE + Sex, InceptionV3 + Bilinear and InceptionV3 + Bilinear. + SE + Sex, to identify the DL model with the best classification performance. Subsequently, the results of the top-performing model were compared with those of manual classification. The study findings revealed that the InceptionV3 + Bilinear + SE + Sex model exhibited the best performance, achieving classification accuracies of 96.15% and 90.48% for the training and test set, respectively. Furthermore, based on the InceptionV3 + Bilinear + SE + Sex model, classification accuracies were calculated for four age groups (< 14.0 years, 14.0 years ≤ age < 16.0 years, 16.0 years ≤ age < 18.0 years, ≥ 18.0 years), with notable accuracies of 100% for the age groups 16.0 years ≤ age < 18.0 years and ≥ 18.0 years. The BA classification, utilizing the feature fusion DL network model, holds significant reference value for determining the age of criminal responsibility of adolescents, particularly at the critical legal age boundaries of 14.0, 16.0, and 18.0 years.

PMID:40402226 | DOI:10.1007/s00414-025-03497-z

J Neurooncol. 2025 May 22. doi: 10.1007/s11060-025-05073-2. Online ahead of print.

ABSTRACT

BACKGROUND: Determining whether pediatric patients with low-grade gliomas (pLGGs) have tumor-related epilepsy (GAE) is a crucial aspect of preoperative evaluation. Therefore, we aim to propose an innovative, machine learning- and deep learning-based framework for the rapid and non-invasive preoperative assessment of GAE in pediatric patients using magnetic resonance imaging (MRI).

METHODS: In this study, we propose a novel radiomics-based approach that integrates tumor and peritumoral features extracted from preoperative multiparametric MRI scans to accurately and non-invasively predict the occurrence of tumor-related epilepsy in pediatric patients.

RESULTS: Our study developed a multimodal MRI radiomics model to predict epilepsy in pLGGs patients, achieving an AUC of 0.969. The integration of multi-sequence MRI data significantly improved predictive performance, with Stochastic Gradient Descent (SGD) classifier showing robust results (sensitivity: 0.882, specificity: 0.956).

CONCLUSION: Our model can accurately predict whether pLGGs patients have tumor-related epilepsy, which could guide surgical decision-making. Future studies should focus on similarly standardized preoperative evaluations in pediatric epilepsy centers to increase training data and enhance the generalizability of the model.

PMID:40402200 | DOI:10.1007/s11060-025-05073-2