Laryngoscope. 2025 May 15. doi: 10.1002/lary.32246. Online ahead of print.

ABSTRACT

OBJECTIVE: Traditional evaluations of surgical skills in otolaryngology rely heavily on subjective assessments, which are prone to variability and bias. This study aims to examine advancements in artificial intelligence (AI) applications for surgical skills evaluation with a focus on their potential to enhance otolaryngology education.

DATA SOURCES: A systematic search of MEDLINE, Embase, Cochrane Database of Systematic Reviews, and Google Scholar was conducted using search terms related to AI and surgical skills evaluation.

REVIEW METHODS: A structured review of the literature up to November 2024 was performed. Two independent reviewers identified and analyzed relevant studies. Reference lists of selected articles were also screened to ensure comprehensiveness.

RESULTS: A total of 34 studies met inclusion criteria. Of these, 56% (19/34) evaluated basic surgical tasks, such as hand-tying, open suturing, and robotic or laparoscopic procedures using bench-top simulators, while 44% (15/34) focused on specific surgical procedures across specialties, including otolaryngology (mastoidectomy, septoplasty, endoscopic sinus surgery, endoscopic carotid injury management), neurosurgery, urology, and general surgery. AI methods applied included deep learning, machine learning, and computer vision techniques. Classification accuracy ranged from 66% to 100% for kinematic, motion, and force data, and from 60% to 96% for video-based analyses of surgical skills.

CONCLUSION: AI-driven assessment tools hold significant promise for otolaryngology surgical education. Automated feedback mechanisms can provide trainees with objective, data-driven insights into their performance, enabling enhanced benchmarking and accelerating learning curves. By adopting AI, otolaryngology has the potential to advance its training methodologies and improve outcomes for both trainees and patients.

LEVEL OF EVIDENCE: N/A.

PMID:40371996 | DOI:10.1002/lary.32246

Stem Cells. 2025 May 15:sxaf031. doi: 10.1093/stmcls/sxaf031. Online ahead of print.

NO ABSTRACT

PMID:40371929 | DOI:10.1093/stmcls/sxaf031

Proteins. 2025 May 15. doi: 10.1002/prot.26835. Online ahead of print.

ABSTRACT

The study aims to design novel therapeutic inhibitors targeting the DHFR protein of Klebsiella pneumoniae. However, challenges like bacterial resistance to peptides and the limitations of computational models in predicting in vivo behavior must be addressed to refine the design process and improve therapeutic efficacy. This study employed deep learning-based bioinformatics techniques to tackle these issues. The study involved retrieving DHFR protein sequences from Klebsiella strains, aligning them to identify conserved regions, and using deep learning models (OmegaFold, ProteinMPNN) to design de novo inhibitors. Cell-penetrating peptide (CPP) motifs were added to enhance delivery, followed by allergenicity and thermal stability assessments. Molecular docking and dynamics simulations evaluated the binding affinity and stability of the inhibitors with DHFR. A conserved 60-residue region was identified, and 60 de novo binders were generated, resulting in 7200 sequences. After allergenicity prediction and stability testing, 10 sequences with melting points near 70°C were shortlisted. Strong binding affinities were observed, especially for complexes 4OR7-1787 and 4OR7-1811, which remained stable in molecular dynamics simulations, indicating their potential as therapeutic agents. This study designed stable de novo peptides with cell-penetrating properties and strong binding affinity to DHFR. Future steps include in vitro validation to assess their effectiveness in inhibiting DHFR, followed by in vivo studies to evaluate their therapeutic potential and stability. These peptides offer a promising strategy against Klebsiella pneumoniae infections, providing potential alternatives to current antibiotics. Experimental validation will be key to assessing their clinical relevance.

PMID:40371895 | DOI:10.1002/prot.26835

Brain Connect. 2025 May 15. doi: 10.1089/brain.2024.0084. Online ahead of print.

ABSTRACT

Objectives: Neurofeedback (NF) based on brain-computer interface (BCI) is an important direction in adjunctive interventions for post-traumatic stress disorder (PTSD). However, existing research lacks comprehensive methodologies and experimental designs. There are concerns in the field regarding the effectiveness and mechanistic interpretability of NF, prompting this study to conduct a systematic analysis of primary NF techniques and research outcomes in PTSD modulation. The study aims to explore reasons behind these concerns and propose directions for addressing them. Methods: A search conducted in the Web of Science database up to December 1, 2023, yielded 111 English articles, of which 80 were excluded based on predetermined criteria irrelevant to this study. The remaining 31 original studies were included in the literature review. A checklist was developed to assess the robustness and credibility of these 31 studies. Subsequently, these original studies were classified into electroencephalogram-based NF (EEG-NF) and functional magnetic resonance imaging-based NF (fMRI-NF) based on BCI type. Data regarding target brain regions, target signals, modulation protocols, control group types, assessment methods, data processing strategies, and reported outcomes were extracted and synthesized. Consensus theories from existing research and directions for future improvements in related studies were distilled. Results: Analysis of all included studies revealed that the average sample size of PTSD patients in EEG and fMRI NF studies was 17.4 (SD 7.13) and 14.6 (SD 6.37), respectively. Due to sample and neurofeedback training protocol constraints, 93% of EEG-NF studies and 87.5% of fMRI-NF studies used traditional statistical methods, with minimal utilization of basic machine learning (ML) methods and no studies utilizing deep learning (DL) methods. Apart from approximately 25% of fMRI NF studies supporting exploratory psychoregulatory strategies, the remaining EEG and fMRI studies lacked explicit NF modulation guidance. Only 13% of studies evaluated NF effectiveness methods involving signal classification, decoding during the NF process, and lacking in process monitoring and assessment means. Conclusion: In summary, NF holds promise as an adjunctive intervention technique for PTSD, potentially aiding in symptom alleviation for PTSD patients. However, improvements are necessary in the process evaluation mechanisms for PTSD-NF, clarity in NF modulation guidance, and development of ML/DL methods suitable for PTSD-NF with small sample sizes. To address these challenges, it is crucial to adopt more rigorous methodologies for monitoring NF, and future research should focus on the integration of advanced data analysis techniques to enhance the effectiveness and precision of PTSD-NF interventions. Impact Statement The implications of this study are to address the limited application of Neurofeedback training (NFT) in post-traumatic stress disorder (PTSD) research, where a significant portion of the approaches, foundational research, and conclusions lack consensus. There is a notable absence of retrospective statistical analyses on NFT interventions for PTSD. This study provides a comprehensive statistical analysis and discussion of existing research, offering valuable insights for future studies. The findings hold significance for researchers, clinicians, and practitioners in the field, providing a foundation for informed, evidence-based interventions for PTSD treatment.

PMID:40371570 | DOI:10.1089/brain.2024.0084

Alzheimer Dis Assoc Disord. 2025 May 15. doi: 10.1097/WAD.0000000000000667. Online ahead of print.

ABSTRACT

BACKGROUND: Genetic testing for dementia has drawn public attention in recent years, albeit with concerns on its appropriate use. This study leveraged Twitter data to analyze public perceptions related to genetic testing for dementia.

METHODS: English tweets from January 1, 2010 to April 1, 2023, containing relevant terms, were extracted from Twitter API. A Bidirectional Encoder Representations from Transformers (BERT) model was used with Named Entity Recognition (NER) to identify individual and organizational users. BERT-based topic modeling was applied to identify the themes for relevant source tweets. Topic coherence was assessed through manual inspection, complemented by the Silhouette Coefficient. Manual thematic analysis, following Braun and Clarke's approach, refined the topics and themes.

RESULTS: The analysis of 3045 original/source tweets identified 9 topics (Silhouette Coefficient=0.19), categorized into 3 main themes: (1) opinions on the appropriateness of genetic testing in dementia diagnosis; (2) discussion on the psychosocial impact; (3) discussion on genetic testing's role in Alzheimer's disease treatment and prevention. Theme 1 comprised 90.6% of source tweets, demonstrating prevailing contentions. Tweets in theme 2 were increasingly contributed by organization users over time and included tweets containing misinformation about genetic testing in children. Tweets in theme 3 were increasingly contributed by individual users, possibly suggesting rising public interest in the treatment and prevention of dementia.

CONCLUSION: The study highlighted limited public understanding of the nondeterministic nature of genetic testing for dementia, with concerns about unsupervised direct-to-consumer genetic test marketing, emphasizing the need to counter misinformation and raise public awareness.

PMID:40371554 | DOI:10.1097/WAD.0000000000000667

Cent European J Urol. 2025;78(1):23-39. doi: 10.5173/ceju.2024.0064. Epub 2025 Mar 21.

ABSTRACT

INTRODUCTION: The incidence of prostate cancer is increasing in Poland, particularly due to the aging population. This review explores the potential of deep learning algorithms to accelerate prostate contouring during fusion biopsies, a time-consuming but crucial process for the precise diagnosis and appropriate therapeutic decision-making in prostate cancer. Implementing convolutional neural networks (CNNs) can significantly improve segmentation accuracy in multiparametric magnetic resonance imaging (mpMRI).

MATERIAL AND METHODS: A comprehensive literature review was conducted using PubMed and IEEE Xplore, focusing on open-access studies from the past five years, and following PRISMA 2020 guidelines. The review evaluates the enhancement of prostate contouring and segmentation in MRI for fusion biopsies using CNNs.

RESULTS: The results indicate that CNNs, particularly those utilizing the U-Net architecture, are predominantly selected for advanced medical image analysis. All the reviewed algorithms achieved a Dice similarity coefficient (DSC) above 74%, indicating high precision and effectiveness in automatic prostate segmentation. However, there was significant heterogeneity in the methods used to evaluate segmentation outcomes across different studies.

CONCLUSIONS: This review underscores the need for developing and optimizing segmentation algorithms tailored to the specific needs of urologists performing fusion biopsies. Future research with larger cohorts is recommended to confirm these findings and further enhance the practical application of CNN-based segmentation tools in clinical settings.

PMID:40371421 | PMC:PMC12073522 | DOI:10.5173/ceju.2024.0064

Data Brief. 2025 Apr 22;60:111564. doi: 10.1016/j.dib.2025.111564. eCollection 2025 Jun.

ABSTRACT

In South Asian countries, agriculture is a crucial employment field, and a remarkable number of people depend on it for their livelihood. Crop diseases are a significant threat to sustainable development in the agriculture field. Automated efficient crop disease diagnosis techniques developed with comprehensive field image datasets can play a vital role in preventing diseases at an early stage. Betel leaf is widely consumed in South Asian countries for its nutritional benefits, but to the best of our knowledge, no extensive dataset of betel leaf is available that can play a crucial role in developing accurate disease diagnosis tools. Farmers face a significant economic loss due to betel leaf diseases, and due to the lack of efficient diagnosis tools, the farming of betel leaf has become very difficult day by day. Our motive is to develop a reliable and versatile image dataset of field images that will assist artificial intelligence-based pathology research on betel leaf diseases. This dataset contains healthy leaf images and two common disease images of betel leaf such as leaf rot and leaf spot [1]. Initially, 2,037 betel leaf images were captured in a natural daylight environment from several betel cultivation fields in Bangladesh. Afterward, 10,185 images were generated using image augmentation strategies including flipping, brightness factor, contrast factor, and rotation. This dataset is well-compatible with machine learning and deep learning-based pathology research, as it contains enough image samples for model training, validation, and testing. Moreover, a comparison study is conducted that ensures this dataset fulfills the gap of a reliable and extensive dataset of betel leaf. This comprehensive dataset serves as a crucial resource for researchers in developing efficient computational models for accurate betel leaf disease diagnosis.

PMID:40371167 | PMC:PMC12076791 | DOI:10.1016/j.dib.2025.111564

Front Neurol. 2025 Apr 30;16:1587347. doi: 10.3389/fneur.2025.1587347. eCollection 2025.

ABSTRACT

BACKGROUND: Acute ischemic stroke (AIS) is a major global health threat associated with high rates of disability and mortality, highlighting the need for early prognostic assessment to guide treatment. Currently, there are no reliable methods for the early prediction of poor prognosis in AIS, especially after mechanical thrombectomy. This study aimed to explore the value of radiomics and deep learning based on multimodal magnetic resonance imaging (MRI) in predicting poor prognosis in patients with AIS who underwent mechanical thrombectomy. This study aimed to provide a more accurate and comprehensive tool for stroke prognosis.

METHODS: This study retrospectively analyzed the clinical data and multimodal MRI images of patients with stroke at admission. Logistic regression was employed to identify the risk factors associated with poor prognosis and to construct a clinical model. Radiomics features of the stroke-affected regions were extracted from the patients' baseline multimodal MRI images, and the optimal radiomics features were selected using a least absolute shrinkage and selection operator regression model combined with five-fold cross-validation. The radiomics score was calculated based on the feature weights, and machine learning techniques were applied using a logistic regression classifier to develop the radiomics model. In addition, a deep learning model was devised using ResNet101 and transfer learning. The clinical, radiomics, and deep learning models were integrated to establish a comprehensive multifactorial logistic regression model, termed the CRD (Clinic-Radiomics-Deep Learning) model. The performance of each model in predicting poor prognosis was assessed using receiver operating characteristic (ROC) curve analysis, with the optimal model visualized as a nomogram. A calibration curve was plotted to evaluate the accuracy of nomogram predictions.

RESULTS: A total of 222 patients with AIS were enrolled in this study in a 7:3 ratio, with 155 patients in the training cohort and 67 in the validation cohort. Statistical analysis of clinical data from the training and validation cohorts identified two independent risk factors for poor prognosis: the National Institutes of Health Stroke Scale score at admission and the occurrence of intracerebral hemorrhage. Of the 1,197 radiomic features, 16 were selected to develop the radiomics model. Area under the ROC curve (AUC) analysis of specific indicators demonstrated varying performances across methods and cohorts. In the training cohort, the clinical, radiomics, deep learning, and integrated CRD models achieved AUC values of 0.762, 0.755, 0.689, and 0.834, respectively. In the validation cohort, the clinical model exhibited an AUC of 0.874, the radiomics model achieved an AUC of 0.805, the deep learning model attained an AUC of 0.757, and the CRD model outperformed all models, with an AUC of 0.908. Calibration curves indicated that the CRD model showed exceptional consistency and accuracy in predicting poor prognosis in patients with AIS. Decision curve analysis revealed that the CRD model offered the highest net benefit compared with the clinical, radiomics, and deep learning models.

CONCLUSION: The CRD model based on multimodal MRI demonstrated high diagnostic efficacy and reliability in predicting poor prognosis in patients with AIS who underwent mechanical thrombectomy. This model holds considerable potential for assisting clinicians with risk assessment and decision-making for patients experiencing ischemic stroke.

PMID:40371075 | PMC:PMC12074947 | DOI:10.3389/fneur.2025.1587347

Acute Med Surg. 2025 May 14;12(1):e70065. doi: 10.1002/ams2.70065. eCollection 2025 Jan-Dec.

NO ABSTRACT

PMID:40370970 | PMC:PMC12077104 | DOI:10.1002/ams2.70065

BJR Open. 2025 Apr 29;7(1):tzaf008. doi: 10.1093/bjro/tzaf008. eCollection 2025 Jan.

ABSTRACT

OBJECTIVES: Artificial intelligence (AI) deep learning algorithms trained on non-contrast CT scans effectively detect and quantify acute COVID-19 lung involvement. Our study explored whether radiological contrast affects the accuracy of AI-measured lung opacities, potentially impacting clinical decisions. We compared lung opacity measurements from AI software with visual assessments by radiologists using CT pulmonary angiography (CTPA) images of early-stage COVID-19 patients.

METHODS: This prospective single-centre study included 18 COVID-19 patients who underwent CTPA due to suspected pulmonary embolism. Patient demographics, clinical data, and 30-day and 90-day mortality were recorded. AI tool (Pulmonary Density Plug-in, AI-Rad Companion Chest CT, SyngoVia; Siemens Healthineers, Forchheim, Germany) was used to estimate the quantity of opacities. Visual quantitative assessments were performed independently by 2 radiologists.

RESULTS: There was a positive correlation between radiologist estimations (r 2 = 0.57) and between the AI data and the mean of the radiologists' estimations (r 2 = 0.70). Bland-Altman plot analysis showed a mean bias of +3.06% between radiologists and -1.32% between the mean radiologist vs AI, with no outliers outside 2×SD for respective comparison.

The AI protocol facilitated a quantitative assessment of lung opacities and showed a strong correlation with data obtained from 2 independent radiologists, demonstrating its potential as a complementary tool in clinical practice.

CONCLUSION: In assessing COVID-19 lung opacities in CTPA images, AI tools trained on non-contrast images, provide comparable results to visual assessments by radiologists.

ADVANCES IN KNOWLEDGE: The Pulmonary Density Plug-in enables quantitative analysis of lung opacities in COVID-19 patients using contrast-enhanced CT images, potentially streamlining clinical workflows and supporting timely decision-making.

PMID:40370862 | PMC:PMC12077292 | DOI:10.1093/bjro/tzaf008

Front Artif Intell. 2025 Apr 30;8:1529814. doi: 10.3389/frai.2025.1529814. eCollection 2025.

ABSTRACT

BACKGROUND: Wireless Capsule Endoscopy (WCE) enables non-invasive imaging of the gastrointestinal tract but generates vast video data, making real-time and accurate abnormality detection challenging. Traditional detection methods struggle with uncontrolled illumination, complex textures, and high-speed processing demands.

METHODS: This study presents a novel approach using Real-Time Detection Transformer (RT-DETR), a transformer-based object detection model, specifically optimized for WCE video analysis. The model captures contextual information between frames and handles variable image conditions. It was evaluated using the Kvasir-Capsule dataset, with performance assessed across three RT-DETR variants: Small (S), Medium (M), and X-Large (X).

RESULTS: RT-DETR-X achieved the highest detection precision. RT-DETR-M offered a practical trade-off between accuracy and speed, while RT-DETR-S processed frames at 270 FPS, enabling real-time performance. All three models demonstrated improved detection accuracy and computational efficiency compared to baseline methods.

DISCUSSION: The RT-DETR framework significantly enhances precision and real-time performance in gastrointestinal abnormality detection using WCE. Its clinical potential lies in supporting faster and more accurate diagnosis. Future work will focus on further optimization and deployment in endoscopic video analysis systems.

PMID:40370809 | PMC:PMC12075415 | DOI:10.3389/frai.2025.1529814

Front Genet. 2025 Apr 30;16:1583081. doi: 10.3389/fgene.2025.1583081. eCollection 2025.

ABSTRACT

BACKGROUND AND OBJECTIVE: Gene expression analysis plays a critical role in lung cancer research, offering molecular feature-based diagnostic insights that are particularly effective in distinguishing lung cancer subtypes. However, the high dimensionality and inherent imbalance of gene expression data create significant challenges for accurate diagnosis. This study aims to address these challenges by proposing an innovative deep learning-based method for predicting lung cancer subtypes.

METHODS: We propose a method called Exo-LCClassifier, which integrates feature selection, one-dimensional convolutional neural networks (1D CNN), and an improved Wasserstein Generative Adversarial Network (WGAN). First, differential gene expression analysis was performed using DESeq2 to identify significantly expressed genes from both normal and tumor tissues. Next, the enhanced WGAN was applied to augment the dataset, addressing the issue of sample imbalance and increasing the diversity of effective samples. Finally, a 1D CNN was used to classify the balanced dataset, thereby improving the model's diagnostic accuracy.

RESULTS: The proposed method was evaluated using five-fold cross-validation, achieving an average accuracy of 0.9766 ± 0.0070, precision of 0.9762 ± 0.0101, recall of 0.9827 ± 0.0050, and F1-score of 0.9793 ± 0.0068. On an external GEO lung cancer dataset, it also showed strong performance with an accuracy of 0.9588, precision of 0.9558, recall of 0.9678, and F1-score of 0.9616.

CONCLUSION: This study addresses the critical challenge of imbalanced learning in lung cancer gene expression analysis through an innovative computational framework. Our solution integrates three advanced techniques: (1) DESeq2 for differential expression analysis, (2) WGAN for data augmentation, and (3) 1D CNN for feature learning and classification. The source codes are publicly available at: https://github.com/lanlinxxs/Exo-classifier.

PMID:40370696 | PMC:PMC12075553 | DOI:10.3389/fgene.2025.1583081

Proc SPIE Int Soc Opt Eng. 2025 Feb;13405:134050L. doi: 10.1117/12.3047283. Epub 2025 Apr 8.

ABSTRACT

Ultra-high-resolution (UHR) photon-counting detector (PCD) CT offers superior spatial resolution compared to conventional CT, benefiting various clinical areas. However, the UHR resolution also significantly increases image noise, which can limit its clinical adoption in areas such as cardiac CT. In clinical practice, this image noise varies substantially across imaging conditions, such as different diagnostic tasks, patient characteristics (e.g., size), scan protocols, and image reconstruction settings. To address these challenges and provide the full potential of PCD-CT for optimal clinical performance, a convolutional neural network (CNN) denoising algorithm was developed, optimized, and tailored to each specific set of conditions. The algorithm's effectiveness in reducing noise and its impact on coronary artery stenosis quantification across different patient size categories (small: water equivalent diameter <300 mm, medium: 300-320 mm, and large: >320 mm) were objectively assessed. Reconstruction kernels at different sharpness, from Bv60 to Bv76, were investigated to determine optimal settings for each patient size regarding image quality and quantitative assessment of coronary stenosis (in terms of percent diameter stenosis). Our findings indicate that for patients with a water equivalent diameter less than 320 mm, CNN-denoised Bv72 images provide optimal image quality, less blooming artifact, and reduced percent diameter stenosis compared to routine images, while for patients with water equivalent diameter over 320 mm, CNN-denoised Bv60 images are preferable. Quantitatively, the CNN reduces noise-by 85% compared to the input images and 53% compared to commercial iterative reconstructions at strength 4 (QIR4)-while maintaining high spatial resolution and a natural noise texture. Moreover, it enhances stenosis quantification by reducing the percent diameter stenosis measurement by 52% relative to the input and 24% relative to QIR4. These improvements demonstrate the capability of CNN denoising in UHR PCD-CT to enhance image quality and quantitative assessment of coronary artery disease in a manner that is adaptive to patient characteristics and imaging conditions.

PMID:40370652 | PMC:PMC12076256 | DOI:10.1117/12.3047283

Nature. 2025 May 14. doi: 10.1038/s41586-025-09010-1. Online ahead of print.

ABSTRACT

The human cerebral cortex is composed of six layers and dozens of areas that are molecularly and structurally distinct1-4. Although single-cell transcriptomic studies have advanced the molecular characterization of human cortical development, a substantial gap exists owing to the loss of spatial context during cell dissociation5-8. Here we used multiplexed error-robust fluorescence in situ hybridization (MERFISH)9, augmented with deep-learning-based nucleus segmentation, to examine the molecular, cellular and cytoarchitectural development of the human fetal cortex with spatially resolved single-cell resolution. Our extensive spatial atlas, encompassing more than 18 million single cells, spans eight cortical areas across seven developmental time points. We uncovered the early establishment of the six-layer structure, identifiable by the laminar distribution of excitatory neuron subtypes, 3 months before the emergence of cytoarchitectural layers. Notably, we discovered two distinct modes of cortical areal specification during mid-gestation: (1) a continuous, gradual transition observed across most cortical areas along the anterior-posterior axis and (2) a discrete, abrupt boundary specifically identified between the primary (V1) and secondary (V2) visual cortices as early as gestational week 20. This sharp binary transition in V1-V2 neuronal subtypes challenges the notion that mid-gestation cortical arealization involves only gradient-like transitions6,10. Furthermore, integrating single-nucleus RNA sequencing with MERFISH revealed an early upregulation of synaptogenesis in V1-specific layer 4 neurons. Collectively, our findings underscore the crucial role of spatial relationships in determining the molecular specification of cortical layers and areas. This study establishes a spatially resolved single-cell analysis paradigm and paves the way for the construction of a comprehensive developmental atlas of the human brain.

PMID:40369074 | DOI:10.1038/s41586-025-09010-1

Sci Rep. 2025 May 15;15(1):16832. doi: 10.1038/s41598-025-01744-2.

ABSTRACT

PCOS (Poly-Cystic Ovary Syndrome) is a multifaceted disorder that often affects the ovarian morphology of women of their reproductive age, resulting in the development of numerous cysts on the ovaries. Ultrasound imaging typically diagnoses PCOS, which helps clinicians assess the size, shape, and existence of cysts in the ovaries. Nevertheless, manual ultrasound image analysis is often challenging and time-consuming, resulting in inter-observer variability. To effectively treat PCOS and prevent its long-term effects, prompt and accurate diagnosis is crucial. In such cases, a prediction model based on deep learning can help physicians by streamlining the diagnosis procedure, reducing time and potential errors. This article proposes a novel integrated approach, QEI-SAM (Quality Enhanced Image - Segment Anything Model), for enhancing image quality and ovarian cyst segmentation for accurate prediction. GAN (Generative Adversarial Networks) and CNN (Convolutional Neural Networks) are the most recent cutting-edge innovations that have supported the system in attaining the expected result. The proposed QEI-SAM model used Enhanced Super Resolution Generative Adversarial Networks (ESRGAN) for image enhancement to increase the resolution, sharpening the edges and restoring the finer structure of the ultrasound ovary images and achieved a better SSIM of 0.938, PSNR value of 38.60 and LPIPS value of 0.0859. Then, it incorporates the Segment Anything Model (SAM) to segment ovarian cysts and achieve the highest Dice coefficient of 0.9501 and IoU score of 0.9050. Furthermore, Convolutional Neural Network - ResNet 50, ResNet 101, VGG 16, VGG 19, AlexNet and Inception v3 have been implemented to diagnose PCOS promptly. Finally, VGG 19 has achieved the highest accuracy of 99.31%.

PMID:40369044 | DOI:10.1038/s41598-025-01744-2

Sci Rep. 2025 May 14;15(1):16804. doi: 10.1038/s41598-025-00890-x.

ABSTRACT

Deep learning, particularly convolutional neural networks (CNNs), has proven valuable for brain tumor classification, aiding diagnostic and therapeutic decisions in medical imaging. Despite their accuracy, these models are vulnerable to adversarial attacks, compromising their reliability in clinical settings. In this research, we utilized a VGG16-based CNN model to classify brain tumors, achieving 96% accuracy on clean magnetic resonance imaging (MRI) data. To assess robustness, we exposed the model to Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks, which reduced accuracy to 32% and 13%, respectively. We then applied a multi-layered defense strategy, including adversarial training with FGSM and PGD examples and feature squeezing techniques such as bit-depth reduction and Gaussian blurring. This approach improved model resilience, achieving 54% accuracy on FGSM and 47% on PGD adversarial examples. Our results highlight the importance of proactive defense strategies for maintaining the reliability of AI in medical imaging under adversarial conditions.

PMID:40369011 | DOI:10.1038/s41598-025-00890-x

J Cheminform. 2025 May 14;17(1):74. doi: 10.1186/s13321-025-01012-5.

ABSTRACT

Toxicity is a critical hurdle in drug development, often causing the late-stage failure of promising compounds. Existing computational prediction models often focus on single-organ toxicity. However, avoiding toxicity of an organ, such as reducing gastrointestinal side effects, may inadvertently lead to toxicity in another organ, as seen in the real case of rofecoxib, which was withdrawn due to increased cardiovascular risks. Thus, simultaneous prediction of multi-organ toxicity is a desirable but challenging task. The main challenges are (1) the variability of substructures that contribute to toxicity of different organs, (2) insufficient power of molecular representations in diverse perspectives, and (3) explainability of prediction results especially in terms of substructures or potential toxicophores. To address these challenges with multiple strategies, we developed FATE-Tox, a novel multi-view deep learning framework for multi-organ toxicity prediction. For variability of substructures, we used three fragmentation methods such as BRICS, Bemis-Murcko scaffolds, and RDKit Functional Groups to formulate fragment-level graphs so that diverse substructures can be used to identify toxicity for different organs. For insufficient power of molecular representations, we used molecular representations in both 2D and 3D perspectives. For explainability, our fragment attention transformer identifies potential 3D toxicophores using attention coefficients. Scientific contribution: Our framework achieved significant improvements in prediction performance, with up to 3.01% gains over prior baseline methods on toxicity benchmark datasets from MoleculeNet (BBBP, SIDER, ClinTox) and TDC (DILI, Skin Reaction, Carcinogens, and hERG), while the multi-task learning approach further enhanced performance by up to 1.44% compared to the single-task learning framework that had already surpassed these baselines. Additionally, attention visualization aligning with literature contributes to greater transparency in predictive modeling. Our approach has the potential to provide scientists and clinicians with a more interpretable and clinically meaningful tool to assess systemic toxicity, ultimately supporting safer and more informed drug development processes.

PMID:40369624 | DOI:10.1186/s13321-025-01012-5

Breast Cancer Res. 2025 May 14;27(1):80. doi: 10.1186/s13058-025-02033-6.

ABSTRACT

BACKGROUND: This study aimed to develop a BI-RADS network (DL-UM) via integrating ultrasound (US) and mammography (MG) images and explore its performance in improving breast lesion diagnosis and management when collaborating with radiologists, particularly in cases with discordant US and MG Breast Imaging Reporting and Data System (BI-RADS) classifications.

METHODS: We retrospectively collected image data from 1283 women with breast lesions who underwent both US and MG within one month at two medical centres and categorised them into concordant and discordant BI-RADS classification subgroups. We developed a DL-UM network via integrating US and MG images, and DL networks using US (DL-U) or MG (DL-M) alone, respectively. The performance of DL-UM network for breast lesion diagnosis was evaluated using ROC curves and compared to DL-U and DL-M networks in the external testing dataset. The diagnostic performance of radiologists with different levels of experience under the assistance of DL-UM network was also evaluated.

RESULTS: In the external testing dataset, DL-UM outperformed DL-M in sensitivity (0.962 vs. 0.833, P = 0.016) and DL-U in specificity (0.667 vs. 0.526, P = 0.030), respectively. In the discordant BI-RADS classification subgroup, DL-UM achieved an AUC of 0.910. The diagnostic performance of four radiologists improved when collaborating with the DL-UM network, with AUCs increased from 0.674-0.772 to 0.889-0.910, specificities from 52.1%-75.0 to 81.3-87.5% and reducing unnecessary biopsies by 16.1%-24.6%, particularly for junior radiologists. Meanwhile, DL-UM outputs and heatmaps enhanced radiologists' trust and improved interobserver agreement between US and MG, with weighted kappa increased from 0.048 to 0.713 (P < 0.05).

CONCLUSIONS: The DL-UM network, integrating complementary US and MG features, assisted radiologists in improving breast lesion diagnosis and management, potentially reducing unnecessary biopsies.

PMID:40369585 | DOI:10.1186/s13058-025-02033-6

BMC Med Inform Decis Mak. 2025 May 14;25(1):184. doi: 10.1186/s12911-025-03011-w.

ABSTRACT

Lung cancer is one of the most prevalent diseases affecting people and is a main factor in the rising death rate. Recently, Machine Learning (ML) and Deep Learning (DL) techniques have been utilized to detect and classify various types of cancer, including lung cancer. In this research, a DL model, specifically a Convolutional Neural Network (CNN), is proposed to classify lung cancer stages for two types of lung cancer (LUAD and LUSC) using a gene dataset. Evaluating and validating the performance of the proposed model required addressing some common challenges in gene datasets, such as class imbalance and overfitting, due to the low number of samples and the high number of features. These issues were mitigated by deeply analyzing the gene dataset and lung cancer stages from a medical perspective, along with extensive research and experiments. As a result, the optimized CNN model using F-test feature selection method, achieved high classification accuracies of approximately 93.94% for LUAD and 88.42% for LUSC.

PMID:40369502 | DOI:10.1186/s12911-025-03011-w

Behav Res Methods. 2025 May 14;57(6):172. doi: 10.3758/s13428-025-02679-2.

ABSTRACT

Driver gaze estimation is important for various driver gaze applications such as building advanced driving assistance systems and understanding driver gaze behavior. Gaze estimation in terms of gaze zone classification requires large-scale labeled data for supervised machine learning and deep learning-based models. In this study, we collected a driver gaze dataset and annotated it using three annotation approaches - manual annotation, Speak2Label, and moving pointer-based annotation. Moving pointer-based annotation was introduced as a new data annotation approach inspired by screen-based gaze data collection. For each data collection approach, ground truth labels were obtained using an eye tracker. The proposed moving pointer-based approach was found to achieve higher accuracy compared to the other two approaches. Due to the lower accuracy of manual annotation and the Speak2Label method, we performed a detailed analysis of these two annotation approaches to understand the reasons for the misclassification. A confusion matrix was also plotted to compare the manually assigned gaze labels with the ground truth labels. This was followed by misclassification analysis, two-sample t-test-based analysis to understand if head pose and pupil position of driver influence the misclassification by the annotators. In Speak2Label, misclassification was observed due to a lag between the speech and gaze time series, which can be visualized in the graph and cross-correlation analysis were done to compute the maximum lag between the two time series. Finally, we created a benchmark Eye Tracker-based Driver Gaze Dataset (ET-DGaze) that consists of the driver's face images and corresponding gaze labels obtained from the eye tracker.

PMID:40369353 | DOI:10.3758/s13428-025-02679-2