PLoS One. 2025 Jun 2;20(6):e0324496. doi: 10.1371/journal.pone.0324496. eCollection 2025.

ABSTRACT

Quantitative MRI (qMRI) requires the acquisition of multiple images with parameter changes, resulting in longer measurement times than conventional imaging. Deep learning (DL) for image reconstruction has shown a significant reduction in acquisition time and improved image quality. In qMRI, where the image contrast varies between sequences, preparing large, fully-sampled (FS) datasets is challenging. Recently, methods that do not require FS data such as self-supervised learning (SSL) and zero-shot self-supervised learning (ZSSSL) have been proposed. Another challenge is the large GPU memory requirement for DL-based qMRI image reconstruction, owing to the simultaneous processing of multiple contrast images. In this context, Kellman et al. proposed memory-efficient learning (MEL) to save the GPU memory. This study evaluated SSL and ZSSSL frameworks with MEL to accelerate qMRI. Three experiments were conducted using the following sequences: 2D T2 mapping/MSME (Experiment 1), 3D T1 mapping/VFA-SPGR (Experiment 2), and 3D T2 mapping/DESS (Experiment 3). Each experiment used the undersampled k-space data under acceleration factors of 4, 8, and 12. The reconstructed maps were evaluated using quantitative metrics. In this study, we performed three qMRI reconstruction measurements and compared the performance of the SL- and GT-free learning methods, SSL and ZSSSL. Overall, the performances of SSL and ZSSSL were only slightly inferior to those of SL, even under high AF conditions. The quantitative errors in diagnostically important tissues (WM, GM, and meniscus) were small, demonstrating that SL and ZSSSL performed comparably. Additionally, by incorporating a GPU memory-saving implementation, we demonstrated that the network can operate on a GPU with a small memory (<8GB) with minimal speed reduction. This study demonstrates the effectiveness of memory-efficient GT-free learning methods using MEL to accelerate qMRI.

PMID:40455714 | DOI:10.1371/journal.pone.0324496

J Appl Clin Med Phys. 2025 Jun 2:e70124. doi: 10.1002/acm2.70124. Online ahead of print.

ABSTRACT

PURPOSE: This study aims to develop a CycleGAN based denoising model to enhance the quality of low-dose PET (LDPET) images, making them as close as possible to standard-dose PET (SDPET) images.

METHODS: Using a Philips Vereos PET/CT system, whole-body PET images of fluorine-18 fluorodeoxyglucose (18F-FDG) were acquired from 37 patients to facilitate the development of the UR-CycleGAN model. In this model, low-dose data were simulated by reconstructing PET images with a 30-s acquisition time, while standard-dose data were reconstructed from a 2.5-min acquisition. The network was trained in a supervised manner on 13 210 pairs of PET images, and the quality of the images was objectively evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).

RESULTS: Compared to simulated low-dose data, the denoised PET images generated by our model showed significant improvement, with a clear trend toward SDPET image quality.

CONCLUSION: The proposed method reduces acquisition time by 80% compared to standard-dose imaging, while achieving image quality close to SDPET images. It also enhances visual detail fidelity, demonstrating the feasibility and practical utility of the model for significantly reducing imaging time while maintaining high image quality.

PMID:40455649 | DOI:10.1002/acm2.70124

J Chem Inf Model. 2025 Jun 2. doi: 10.1021/acs.jcim.5c00556. Online ahead of print.

ABSTRACT

The black-box nature of deep learning has increasingly drawn attention to the reliability and uncertainty of predictive models. Currently, several uncertainty quantification (UQ) methods have been proposed and successfully applied in the fields of molecules and proteins, effectively improving model prediction quality and interpretability. Protein-RNA binding represents a fundamental aspect of protein research. Accurate prediction of binding sites and ensuring the reliability of such predictions are crucial for various scientific endeavors. However, many of the existing computational methods have a single feature extraction and lack of UQ. To address these, we propose MGCA (multiscale graph convolutional networks, convolutional neural networks and attention) to better capture local and global information and achieve competitive results in predicting protein-RNA binding sites. Moreover, we launch a UQ study based on MGCA and five prevalent models to verify the robustness of the results. Specifically, we introduce the Expected Calibration Error (ECE) to assess the uncertainty of the models. Additionally, a novel split-bins screening method is proposed based on the ECE, aiming to investigate the practical impact of reducing uncertainty on the models. Finally, temperature scaling (TS) is used to calibrate model uncertainty without changing performance. Results show that the split-bins screening method reduces false positives (FP), and TS significantly decreases the model ECE. The split-bins screening method combined with TS can further reduce FP and improve precision. Our findings demonstrate that TS effectively reduces uncertainty in protein-RNA binding site prediction, and minimizing model uncertainty enhances prediction quality. The data and code can be available at https://github.com/trustcm/UQ-TS-Split-bins-RBP.

PMID:40455481 | DOI:10.1021/acs.jcim.5c00556

Interdiscip Sci. 2025 Jun 2. doi: 10.1007/s12539-025-00720-8. Online ahead of print.

ABSTRACT

Accurate pulmonary nodule detection in CT imaging remains challenging due to fragmented feature integration in conventional deep learning models. This paper proposes SPCF-YOLO, a real-time detection framework that synergizes hierarchical feature fusion with anatomical context modeling. First, the space-to-depth convolution (SPDConv) module preserves fine-grained features in low-resolution images through spatial dimension reorganization. Second, the shared feature pyramid convolution (SFPConv) module is designed to dynamically extract multi-scale contextual information using multi-dilation-rate convolutional layers. Incorporating a small object detection layer aims to improve sensitivity to small nodules. This is achieved in combination with the improved pyramid squeeze attention (PSA) module and the improved contextual transformer (CoTB) module, which enhance global channel dependencies and reduce feature loss. The model achieves 82.8% mean average precision (mAP) and 82.9% F1 score on LUNA16 at 151 frames per second (representing improvements of 17.5% and 82.9% over YOLOv8 respectively), demonstrating real-time clinical viability. Cross-modality validation on SIIM-COVID-19 shows 1.5% improvement, confirming robust generalization.

PMID:40455403 | DOI:10.1007/s12539-025-00720-8

Interdiscip Sci. 2025 Jun 2. doi: 10.1007/s12539-025-00713-7. Online ahead of print.

ABSTRACT

With the advantages of reducing biochemical experiments and enabling the rapid screening of potential druggable compounds, accurate computational methods are essential for predicting Drug-Target affinity (DTA). Current deep learning-based DTA prediction methods predominantly concentrate on single-modal information from drugs or targets. In this article, we propose a new multi-modal DTA prediction method, MGSDTA, to integrate graph features and sequence features of drug molecules and target proteins. We extract features from the drug molecular graphs and target protein graphs, meanwhile, we extract sequence features using continuous embeddings generated by advanced self-supervised pre-trained models, Mol2vec and ProtVec, for drug substructures and target subsequences respectively. Finally, they are integrated with a weighted fusion module for DTA prediction. Experiments on benchmark datasets indicate that the performance of MGSDTA exceeds single-modal methods based solely on sequences or graphs.

PMID:40455402 | DOI:10.1007/s12539-025-00713-7

Interdiscip Sci. 2025 Jun 2. doi: 10.1007/s12539-025-00716-4. Online ahead of print.

ABSTRACT

Non-coding RNAs (ncRNAs) are one of the components of epigenetic mechanisms that regulates gene expression. Studying ncRNA-protein interactions (NPI) can help to explore a wide range of biological features and related diseases. Traditional NPI research methods often require expensive equipment, a lot of time and labor. With the abundant samples accumulated from traditional experiments, remarkable progress has been made in the study of NPI by computational methods. Heterogeneous graph neural network is a deep learning method that synthesizes heterogeneous types of data as well as network topology. In this study, we propose an NPI-HetGNN model for NPI prediction based on heterogeneous graph neural networks. Firstly, initial features are constructed by integrating the sequence properties of ncRNA and protein data as well as the topology of heterogeneous connections. Then, the multilevel homogeneous subgraph is obtained and its semantic information is aggregated by metapath walking. At the same time, the homogeneous node information is fused within the subgraph metapath. To enhance feature extraction ability of the network, an energy-constrained self-attention module is introduced. Due to the lack of wet lab validation conditions, this study adopts computational verification. The performance of the NPI-HetGNN model on four benchmark datasets is experimentally verified. Ablation experiments also confirmed the comprehensiveness and validity of our model design. The experimental results show that comparing with six state-of-the-art methods, our NPI-HetGNN achieves very satisfactory results on all four datasets.

PMID:40455400 | DOI:10.1007/s12539-025-00716-4

Leuk Lymphoma. 2025 Jun 2:1-13. doi: 10.1080/10428194.2025.2507198. Online ahead of print.

ABSTRACT

Monoclonal antibody (mAb)-based therapies targeting CD19, CD20, and CD22 have revolutionized B-ALL treatment, offering precision and reduced systemic toxicity by engaging immune mechanisms to eliminate leukemic cells. This review synthesizes literature from PubMed, Web of Science, and ClinicalTrials.gov (2000-2024), focusing on clinical outcomes and resistance mechanisms. Bispecific T-cell engagers (e.g. blinatumomab) and CD22-directed antibody-drug conjugates (e.g. inotuzumab ozogamicin) demonstrate robust efficacy in relapsed/refractory disease. Advances in antibody engineering, such as Fc optimization, nanobodies, and humanization, enhance tumor targeting and therapeutic safety. Persistent challenges include antigen escape, stromal-mediated resistance, and treatment-related toxicities. Combinatorial approaches integrating mAbs with CAR-T cells or checkpoint inhibitors show promise in overcoming resistance pathways. Emerging technologies like artificial intelligence and deep learning are transforming antibody design by predicting epitope binding, enabling de novo protein engineering, and streamlining affinity maturation. These innovations accelerate the development of next-generation therapies, underscoring the evolving potential of precision immunotherapy of B-ALL.

PMID:40455243 | DOI:10.1080/10428194.2025.2507198

Chaos. 2025 Jun 1;35(6):063101. doi: 10.1063/5.0264041.

ABSTRACT

The "bad" Jaulent-Miodek (JM) equation serves to describe the motion of non-viscous shallow water wave packets in a flat-bottomed domain subject to shear forces. The "bad" JM equation exhibits poor properties, characterized by the linear instability of nonlinear waves on the zero-plane background, rendering it challenging to solve through traditional analytical and numerical methods. In this paper, two classic physics-driven deep learning approaches, namely, Physics-Informed Neural Networks (PINN) and Physics and Equality-Constrained Artificial Neural Networks (PECANN), are combined into a two-stage "PINN+PECANN" neural network to address the nonlinear wave evolution on the zero-plane background for the "bad" JM equation. The two-stage "PINN+PECANN" neural network method employs PINN in the first stage to pre-train the neural network, followed by fine-tuning of the network parameters using PECANN in the second stage. This approach not only correctly obtains solutions to the "bad" JM equation but also enhances computational efficiency. Specifically, we present the evolutionary behavior of nonlinear waves for the common initial values of the "bad" JM equation: Gauss wave packets, sech wave packets, and rational wave packets. Furthermore, the nonlinear interactions between two Gauss, sech, rational wave packets are provided. The results in this paper validate the advantages of physics-driven deep learning methods in solving equations with poor properties and open up a new pathway for obtaining unstable solutions of nonlinear equations.

PMID:40455205 | DOI:10.1063/5.0264041

Methods Mol Biol. 2025;2927:195-220. doi: 10.1007/978-1-0716-4546-8_11.

ABSTRACT

The COVID crisis has accelerated the integration of artificial intelligence (AI) in drug discovery and omics research, providing novel avenues to tackle intricate issues in virology research. AI has lately enabled significant breakthroughs in a wide range of biological disciplines, including genetic variant interpretation, protein structure prediction, disease detection, and pharmaceutical creation. It has prominently assumed a pivotal role in virology research, with generative AI at the forefront of innovation. Generative AI (GAI) is a subset of AI that majorly focuses on creating new data or content resembling existing data through learning underlying patterns and relationships. It has revolutionized virology/omics study by generating synthetic data to augment limited datasets, predicting protein structures, identifying gene regulatory networks, and assisting in drug discovery through virtual screening, accelerating advancements in genomics, proteomics, and metabolomics research. This chapter aims to discuss the basic concept of generative models and their current and future scope in virology.

PMID:40455159 | DOI:10.1007/978-1-0716-4546-8_11

Clin Oral Investig. 2025 Jun 2;29(6):328. doi: 10.1007/s00784-025-06410-5.

ABSTRACT

OBJECTIVE: The aim of this study was to develop convolutional neural network (CNN)-based systems to diagnose calculus, plaque, gingival hyperplasia and gingival inflammation in intraoral images from orthodontic patients.

MATERIALS AND METHODS: A dataset of 1,000 lateral and frontal intraoral images from orthodontic patients was used to develop CNN-based models. Periodontology specialists annotated areas of dental calculus, plaque, gingival inflammation, and gingival hyperplasia on the teeth and gingiva. The dataset was divided into training (80%), validation (10%), and test (10%) sets for model development. The YOLOv8 and hybrid U-Net + ResNet50 models were examined. Their performance was evaluated on the basis of accuracy, precision, recall, F1 score, Tversky loss, intersection over union, mean average precision, Dice coefficient, and Cohen's kappa.

RESULTS: The mean classification accuracy was 0.96 for the YOLOv8 model and 0.93 for the U-Net + ResNet50 model. On the basis of the Dice coefficient, the models performed best in detecting gingival hyperplasia (YOLOv8: 0.78, U-Net + ResNet50: 0.79) and worst in detecting dental calculus (YOLOv8:0.48, U-Net + ResNet50:0.53). Cohen's kappa coefficient was highest for classifying gingival hyperplasia (YOLOv8: 0.785, U-Net + ResNet50: 0.790). The precision exceeded 0.72 across all the classifications, with the greatest precision in classifying gingival inflammation.

CONCLUSION: Deep learning-based systems can serve as decision support tools by offering rapid and objective evaluations of dental calculus, plaque, gingival inflammation, and gingival hyperplasia. Nonetheless, the definitive diagnostic conclusion should be based on the clinician's specialized expertise and professional judgment.

CLINICAL RELEVANCE: The integration of CNN-based diagnostic models into clinical workflows has the potential to facilitate early periodontal diagnosis and improve accessibility to periodontal assessments in orthodontic patients.

PMID:40455084 | DOI:10.1007/s00784-025-06410-5

Transl Vis Sci Technol. 2025 Jun 2;14(6):3. doi: 10.1167/tvst.14.6.3.

ABSTRACT

PURPOSE: Standard deep learning (DL) models often suffer significant performance degradation on out-of-distribution (OOD) data, where test data differs from training data, a common challenge in medical imaging due to real-world variations.

METHODS: We propose a unified self-censorship framework as an alternative to the standard DL models for glaucoma classification using deep evidential uncertainty quantification. Our approach detects OOD samples at both the dataset and image levels. Dataset-level self-censorship enables users to accept or reject predictions for an entire new dataset based on model uncertainty, whereas image-level self-censorship refrains from making predictions on individual OOD images rather than risking incorrect classifications. We validated our approach across diverse datasets.

RESULTS: Our dataset-level self-censorship method outperforms the standard DL model in OOD detection, achieving an average 11.93% higher area under the curve (AUC) across 14 OOD datasets. Similarly, our image-level self-censorship model improves glaucoma classification accuracy by an average of 17.22% across 4 external glaucoma datasets against baselines while censoring 28.25% more data.

CONCLUSIONS: Our approach addresses the challenge of generalization in standard DL models for glaucoma classification across diverse datasets by selectively withholding predictions when the model is uncertain. This method reduces misclassification errors compared to state-of-the-art baselines, particularly for OOD cases.

TRANSLATIONAL RELEVANCE: This study introduces a tunable framework that explores the trade-off between prediction accuracy and data retention in glaucoma prediction. By managing uncertainty in model outputs, the approach lays a foundation for future decision support tools aimed at improving the reliability of automated glaucoma diagnosis.

PMID:40455037 | DOI:10.1167/tvst.14.6.3

J Chem Inf Model. 2025 Jun 2. doi: 10.1021/acs.jcim.5c00538. Online ahead of print.

ABSTRACT

Predictive models hold considerable promise in enabling the faster discovery of safer, more efficacious therapeutics. To better understand and improve the performance of small-molecule predictive models for drug discovery, we conduct multiple experiments with deep learning and traditional machine learning approaches, leveraging our large internal data sets as well as publicly available data sets. The experiments include assessing performance on random, temporal, and reverse-temporal data ablation tasks as well as tasks testing model extrapolation to different property spaces. We identify factors that contribute to the higher performance of predictive models built using graph neural networks compared to traditional methods such as XGBoost and random forest. These insights were successfully used to develop a scaling relationship that explains 81% of the variance in model performance across various assays and data regimes. This relationship can be used to estimate the performance of models for ADMET (absorption, distribution, metabolism, excretion, and toxicity) end points, as well as for drug discovery assay data more broadly. The findings offer guidance for further improving model performance in drug discovery.

PMID:40454949 | DOI:10.1021/acs.jcim.5c00538

J Proteome Res. 2025 Jun 2. doi: 10.1021/acs.jproteome.5c00063. Online ahead of print.

ABSTRACT

A fundamental challenge in mass spectrometry-based proteomics is determining which peptide generated a given MS2 spectrum. Peptide sequencing typically relies on matching spectra against a known sequence database, which in some applications is not available. Deep learning-based de novo sequencing can address this limitation by directly predicting peptide sequences from MS2 data. We have seen the application of the transformer architecture to de novo sequencing produce state-of-the-art results on the so-called nine-species benchmark. In this study, we propose an improved transformer encoder inspired by the heuristics used in the manual interpretation of spectra. We modify the attention mechanism with a learned bias based on pairwise mass differences, termed Pairwise Attention (PA). Adding PA improves average peptide precision at 100% coverage by 12.7% (5.9 percentage points) over our base transformer on the original nine-species benchmark. We have also achieved a 7.4% increase over the previously published model Casanovo. Our MS2 encoding strategy is largely orthogonal to other transformer-based models encoding MS2 spectra, enabling straightforward integration into existing deep-learning approaches. Our results show that integrating domain-specific knowledge into transformers boosts de novo sequencing performance.

PMID:40454436 | DOI:10.1021/acs.jproteome.5c00063

Nano Lett. 2025 Jun 2. doi: 10.1021/acs.nanolett.5c01460. Online ahead of print.

ABSTRACT

Moiré patterns in van der Waals bilayer materials complicate the analysis of atomic-resolution images, hindering the atomic-scale insight typically attainable with scanning transmission electron microscopy. Here, we report a method to detect the positions and identities of atoms in each of the individual layers that compose twisted bilayer heterostructures. We developed a deep learning model, Gomb-Net, which identifies the coordinates and atomic species in each layer, deconvoluting the moiré pattern. This enables layer-specific mapping of atomic positions and dopant distributions, unlike other commonly used segmentation models which struggle with moiré-induced complexity. Using this approach, we explored the Se atom substitutional site distribution in a twisted fractional Janus WS2-WS2(1-x)Se2x heterostructure and found that layer-specific implantation sites are unaffected by the moiré pattern's local energetic or electronic modulation. This advancement enables atom identification within material regimes where it was not possible before, opening new insights into previously inaccessible material physics.

PMID:40454431 | DOI:10.1021/acs.nanolett.5c01460

Biomicrofluidics. 2025 May 29;19(3):031304. doi: 10.1063/5.0252554. eCollection 2025 May.

ABSTRACT

Microfluidic biochips (MBs) are transforming diagnostics, healthcare, and biomedical research. However, their rapid deployment has exposed them to diverse security threats, including structural tampering, material degradation, sample-level interference, and intellectual property (IP) theft, such as counterfeiting, overbuilding, and piracy. This perspective highlights emerging attack vectors and countermeasures aimed at mitigating these risks. Structural attacks, such as stealthy design code modifications, can result in faulty diagnostics. To address this, deep learning -based anomaly detection leverages microstructural changes, including optical changes such as shadows or reflections, to identify and resolve faults. Material-level countermeasures, including mechano-responsive dyes and spectrometric watermarking, safeguard against subtle chemical alterations during fabrication. Sample-level protections, such as molecular barcoding, ensure bio-sample integrity by embedding unique DNA sequences for authentication. At the IP level, techniques like watermarking, physically unclonable functions, fingerprinting, and obfuscation schemes provide robust defenses against reverse engineering and counterfeiting. Together, these approaches offer a multi-layered security framework to protect MBs, ensuring their reliability, safety, and trustworthiness in critical applications.

PMID:40454326 | PMC:PMC12124908 | DOI:10.1063/5.0252554

HSS J. 2025 May 28:15563316251339660. doi: 10.1177/15563316251339660. Online ahead of print.

ABSTRACT

As artificial intelligence (AI) advances in healthcare, encompassing robust applications for the diagnosis and prognostication of musculoskeletal diseases, clinicians must increasingly understand the implications of machine learning and deep learning in their practice. This review article explores computer vision algorithms and patient-specific, multimodal prediction models; provides a simple framework to guide discussion on the limitations of AI model development; and introduces the field of generative AI.

PMID:40454292 | PMC:PMC12119539 | DOI:10.1177/15563316251339660

Biol Methods Protoc. 2025 May 5;10(1):bpaf034. doi: 10.1093/biomethods/bpaf034. eCollection 2025.

ABSTRACT

This study introduces an approach to classifying histopathological images for detecting dysplasia in oral cancer through the fusion of support vector machine (SVM) classifiers trained on deep learning features extracted from InceptionResNet-v2 and vision transformer (ViT) models. The classification of dysplasia, a critical indicator of oral cancer progression, is often complicated by class imbalance, with a higher prevalence of dysplastic lesions compared to non-dysplastic cases. This research addresses this challenge by leveraging the complementary strengths of the two models. The InceptionResNet-v2 model, paired with an SVM classifier, excels in identifying the presence of dysplasia, capturing fine-grained morphological features indicative of the condition. In contrast, the ViT-based SVM demonstrates superior performance in detecting the absence of dysplasia, effectively capturing global contextual information from the images. A fusion strategy was employed to combine these classifiers through class selection: the majority class (presence of dysplasia) was predicted using the InceptionResNet-v2-SVM, while the minority class (absence of dysplasia) was predicted using the ViT-SVM. The fusion approach significantly outperformed individual models and other state-of-the-art methods, achieving superior balanced accuracy, sensitivity, precision, and area under the curve. This demonstrates its ability to handle class imbalance effectively while maintaining high diagnostic accuracy. The results highlight the potential of integrating deep learning feature extraction with SVM classifiers to improve classification performance in complex medical imaging tasks. This study underscores the value of combining complementary classification strategies to address the challenges of class imbalance and improve diagnostic workflows.

PMID:40454251 | PMC:PMC12122209 | DOI:10.1093/biomethods/bpaf034

Asia Pac J Sports Med Arthrosc Rehabil Technol. 2025 May 5;42:1-7. doi: 10.1016/j.asmart.2025.04.005. eCollection 2025 Oct.

ABSTRACT

OBJECTIVE: To evaluate the performance of a code free deep learning (CFDL) application in diagnosing supraspinatus tendon pathologies on shoulder magnetic resonance imaging (MRI) images.

DESIGN: This retrospective cross-sectional study included patients with supraspinatus MRI showing partial or full-thickness tears and tendinosis, with patients having normal findings as the control group. MRI images were processed in the LobeAI application using transfer learning with ResNet-50 V2 for model development. Models were built to differentiate each pathology from normal and full-thickness tears from partial tears.

RESULTS: The ML models developed using the LobeAI application demonstrated the ability to differentiate between normal shoulder MRI images and partial tears, full-thickness tears, and tendinosis with sensitivities of 93.75 %, 100 %, and 100 %, respectively, and specificities of 43.75 %, 62.5 %, and 18.75 %. The model designed to classify partial vs. full-thickness tears achieved an accuracy of 34.38 %. The model incorporating all pathological images compared to normal MRI images exhibited an accuracy of 37.50 % and a weighted F1 score of 0.32.

CONCLUSION: The results of the study suggest that, although CFDL applications may be promising for the initial detection of supraspinatus pathologies, their current iteration has limitations that must be resolved before they can be reliably integrated into clinical practice.

PMID:40454208 | PMC:PMC12124677 | DOI:10.1016/j.asmart.2025.04.005

Med Hypothesis Discov Innov Ophthalmol. 2025 May 10;14(1):255-272. doi: 10.51329/mehdiophthal1517. eCollection 2025 Spring.

ABSTRACT

BACKGROUND: By leveraging the imaging-rich nature of ophthalmology and optometry, artificial intelligence (AI) is rapidly transforming the vision sciences and addressing the global burden of ocular diseases. The ability of AI to analyze complex imaging and clinical data allows unprecedented improvements in diagnosis, management, and patient outcomes. In this narrative review, we explore the current and emerging opportunities of utilizing AI in the vision sciences, critically examine the associated challenges, and discuss the ethical implications of integrating AI into clinical practice.

METHODS: We searched PubMed/MEDLINE and Google Scholar for English-language articles published from January 1, 2005, to March 31, 2025. Studies on AI applications in ophthalmology and optometry, focusing on diagnostic performance, clinical integration, and ethical considerations, were included, irrespective of study design (clinical trials, observational studies, validation studies, systematic reviews, and meta-analyses). Articles not related to the use of AI in vision care were excluded.

RESULTS: AI has achieved high diagnostic accuracy across different ocular domains. In terms of the cornea and anterior segment, AI models have detected keratoconus with sensitivity and accuracy exceeding 98% and 99.6%, respectively, including in subclinical cases, by analyzing Scheimpflug tomography and corneal biomechanics. For cataract surgery, machine learning-based intraocular lens power calculation formulas, such as the Kane and ZEISS AI formulas, reduce refractive errors, achieving mean absolute errors below 0.30 diopters and performing particularly well in highly myopic eyes. AI-based retinal screening systems, such as the EyeArt and IDx-DR, can autonomously detect diabetic retinopathy with sensitivities above 95%, while deep learning models can predict age-related macular degeneration progression with an area under the receiver operating characteristic curve exceeding 0.90. In glaucoma detection, fundus and optical coherence tomography-based AI models have reached pooled sensitivity and specificity exceeding 90%, although performance varies with disease stage and population diversity. AI has also advanced strabismus detection, amblyopia risk prediction, and myopia progression forecasting by using facial analysis and biometric data. Currently, key challenges in implementing AI in ophthalmology include dataset bias, limited external validation, regulatory hurdles, and ethical issues, such as transparency and equitable access.

CONCLUSIONS: AI is rapidly transforming vision sciences by improving diagnostic accuracy, streamlining clinical workflow, and broadening access to quality eye care, particularly in underserved regions. Its integration into ophthalmology and optometry thus holds significant promise for enhancing patient outcomes and optimizing healthcare delivery. However, to harness the transformative potential of AI fully, sustained multidisciplinary collaboration, involving clinicians, data scientists, ethicists, and policymakers, is essential. Rigorous validation processes, transparency in algorithm development, and strong ethical oversight are equally important to mitigate risks such as bias, data misuse, and unequal access. Responsible implementation of AI in the vision sciences is essential to ensure that all populations are served equitably.

PMID:40453785 | PMC:PMC12121673 | DOI:10.51329/mehdiophthal1517

Natl Sci Rev. 2025 May 14;12(6):nwaf188. doi: 10.1093/nsr/nwaf188. eCollection 2025 Jun.

NO ABSTRACT

PMID:40453638 | PMC:PMC12125967 | DOI:10.1093/nsr/nwaf188