Chaos. 2025 Aug 1;35(8):083116. doi: 10.1063/5.0273776.

ABSTRACT

The Fokker-Planck (FP) equation governs the probabilistic response of diffusion processes driven by stochastic differential equations (SDEs). Gaussian mixture models and deep learning solvers are two state-of-the-art methods for solving the FP equation. Although mixture models mostly depend on empirical sampling strategies and predefined Gaussian components, deep learning techniques suffer from inherent interpretability deficits and require excessively large training samples. To address these challenges, we propose an adaptive normalizing flow framework for solving FP equations (ANFFP). Normalizing flows are generative models that produce tractable distributions to approximate the complex target distributions. The ANFFP architecture inherently preserves probabilistic interpretability while enabling efficient exact sampling advantages that significantly enhance its applicability to probabilistic response modeling under small sample conditions. Numerical examples involving one-dimensional, two-dimensional, and four-dimensional SDEs demonstrate the effectiveness of the method. In addition, the computational complexity of the ANFFP method is discussed in more detail. This work provides a new paradigm for solving high-dimensional FP equations with theoretical guarantees and practical scalability.

PMID:40779784 | DOI:10.1063/5.0273776

PLoS Comput Biol. 2025 Aug 8;21(8):e1013345. doi: 10.1371/journal.pcbi.1013345. Online ahead of print.

ABSTRACT

High-resolution posture tracking of C. elegans has applications in genetics, neuroscience, and drug screening. While classic methods can reliably track isolated worms on uniform backgrounds, they fail when worms overlap, coil, or move in complex environments. Model-based tracking and deep learning approaches have addressed these issues to an extent, but there is still significant room for improvement in tracking crawling worms. Here we train a version of the DeepTangle algorithm developed for swimming worms using a combination of data derived from Tierpsy tracker and hand-annotated data for more difficult cases. DeepTangleCrawl (DTC) outperforms existing methods, reducing failure rates and producing more continuous, gap-free worm trajectories that are less likely to be interrupted by collisions between worms or self-intersecting postures (coils). We show that DTC enables the analysis of previously inaccessible behaviours and increases the signal-to-noise ratio in phenotypic screens, even for data that was specifically collected to be compatible with legacy trackers including low worm density and thin bacterial lawns. DTC broadens the applicability of high-throughput worm imaging to more complex behaviours that involve worm-worm interactions and more naturalistic environments including thicker bacterial lawns.

PMID:40779582 | DOI:10.1371/journal.pcbi.1013345

PLoS One. 2025 Aug 8;20(8):e0328867. doi: 10.1371/journal.pone.0328867. eCollection 2025.

ABSTRACT

Metabolic dysfunction-associated steatotic liver disease (MASLD) is the most common cause of chronic liver disease worldwide, affecting over 30% of the global general population. Its progressive nature and association with other chronic diseases makes early diagnosis important. MRI Proton Density Fat Fraction (PDFF) is the most accurate noninvasive method for quantitatively assessing liver fat but is expensive and has limited availability; accurately quantifying liver fat from more accessible and affordable imaging could potentially improve patient care. This proof-of-concept study explores the feasibility of inferring liver MRI-PDFF values from contrast-enhanced computed tomography (CECT) using deep learning. In this retrospective, cross-sectional study, we analyzed data from living liver donor candidates who had concurrent CECT and MRI-PDFF as part of their pre-surgical workup between April 2021 and October 2022. Manual MRI-PDFF analysis was performed following a standard of clinical care protocol and used as ground truth. After liver segmentation and registration, a deep neural network (DNN) with 3D U-Net architecture was trained using CECT images as single channel input and the concurrent MRI-PDFF images as single channel output. We evaluated performance using mean absolute error (MAE) and root mean squared error (RMSE), and mean errors (defined as the mean difference of results of comparator groups), with 95% confidence intervals (CIs). We used Kappa statistics and Bland-Altman plots to assess agreement between DNN-predicted PDFF and ground truth steatosis grades and PDFF values, respectively. The final study cohort was of 94 patients, mean PDFF = 3.8%, range 0.2-22.3%. When comparing ground truth to segmented reference (MRI-PDFF), our model had an MAE of 0.56, an RMSE of 0.77, and a mean error of 0.06 (-1.75,1.86); when comparing medians of the predicted and reference MRI-PDFF images, our model had an MAE, an RMSE, and a mean error of 2.94, 4.27, and 1.28 (-4.58,7.14), respectively. We found substantial agreement between categorical steatosis grades obtained from DNN-predicted and clinical ground truth PDFF (kappa = 0.75). While its ability to infer exact MRI-PDFF values from CECT images was limited, categorical classification of fat fraction at lower grades was robust, outperforming other prior attempted methods.

PMID:40779568 | DOI:10.1371/journal.pone.0328867

IEEE Trans Med Imaging. 2025 Aug 8;PP. doi: 10.1109/TMI.2025.3596957. Online ahead of print.

ABSTRACT

Computed tomography (CT) is one of the most widely used non-invasive imaging modalities for medical diagnosis. In clinical practice, CT images are usually acquired with large slice thicknesses due to the high cost of memory storage and operation time, resulting in an anisotropic CT volume with much lower inter-slice resolution than in-plane resolution. Since such inconsistent resolution may lead to difficulties in disease diagnosis, deep learning-based volumetric super-resolution methods have been developed to improve inter-slice resolution. Most existing methods conduct single-image super-resolution on the through-plane or synthesize intermediate slices from adjacent slices; however, the anisotropic characteristic of 3D CT volume has not been well explored. In this paper, we propose a novel cross-view texture transfer approach for CT slice interpolation by fully utilizing the anisotropic nature of 3D CT volume. Specifically, we design a unique framework that takes high-resolution in-plane texture details as a reference and transfers them to low-resolution through-plane images. To this end, we introduce a multi-reference non-local attention module that extracts meaningful features for reconstructing through-plane high-frequency details from multiple in-plane images. Through extensive experiments, we demonstrate that our method performs significantly better in CT slice interpolation than existing competing methods on public CT datasets including a real-paired benchmark, verifying the effectiveness of the proposed framework. The source code of this work is available at https://github.com/khuhm/ACVTT.

PMID:40779378 | DOI:10.1109/TMI.2025.3596957

IEEE Trans Med Imaging. 2025 Aug 8;PP. doi: 10.1109/TMI.2025.3597026. Online ahead of print.

ABSTRACT

Choroidal thickness variations serve as critical biomarkers for numerous ophthalmic diseases. Accurate segmentation and quantification of the choroid in optical coherence tomography (OCT) images is essential for clinical diagnosis and disease progression monitoring. Due to the small number of disease types in the public OCT dataset involving changes in choroidal thickness and the lack of a publicly available labeled dataset, we constructed the Xuzhou Municipal Hospital (XZMH)-Choroid dataset. This dataset contains annotated OCT images of normal and eight choroid-related diseases. However, segmentation of the choroid in OCT images remains a formidable challenge due to the confounding factors of blurred boundaries, non-uniform texture, and lesions. To overcome these challenges, we proposed a mixed attention-guided multiscale feature fusion network (MAMFF-Net). This network integrates a Mixed Attention Encoder (MAE) for enhanced fine-grained feature extraction, a deformable multiscale feature fusion path (DMFFP) for adaptive feature integration across lesion deformations, and a multiscale pyramid layer aggregation (MPLA) module for improved contextual representation learning. Through comparative experiments with other deep learning methods, we found that the MAMFF-Net model has better segmentation performance than other deep learning methods (mDice: 97.44, mIoU: 95.11, mAcc: 97.71). Based on the choroidal segmentation implemented in MAMFF-Net, an algorithm for automated choroidal thickness measurement was developed, and the automated measurement results approached the level of senior specialists.

PMID:40779377 | DOI:10.1109/TMI.2025.3597026

IEEE Trans Image Process. 2025 Aug 8;PP. doi: 10.1109/TIP.2025.3595408. Online ahead of print.

ABSTRACT

Autoencoders (AEs) have received extensive attention in hyperspectral anomaly detection (HAD) due to their capability to separate the background from the anomaly based on the reconstruction error. However, the existing AE methods routinely fail to adequately exploit spatial information and may precisely reconstruct anomalies, thereby affecting the detection accuracy. To address these issues, this study proposes a novel Multi-scale Autoencoder Suppression Strategy (MASS). The underlying principle of MASS is to prioritize the reconstruction of background information over anomalies. In the encoding stage, the Local Feature Extractor, which integrates Convolution and Omni-Dimensional Dynamic Convolution (ODConv), is combined with the Global Feature Extractor based on Transformer to effectively extract multi-scale features. Furthermore, a Self-Attention Suppression module (SAS) is devised to diminish the influence of anomalous pixels, enabling the network to focus more intently on the precise reconstruction of the background. During the process of network learning, a mask derived from the test outcomes of each iteration is integrated into the loss function computation, encompassing only the positions with low anomaly scores from the preceding detection round. Experiments on eight datasets demonstrate that the proposed method is significantly superior to several traditional methods and deep learning methods in terms of performance.

PMID:40779374 | DOI:10.1109/TIP.2025.3595408

Clin Transl Oncol. 2025 Aug 8. doi: 10.1007/s12094-025-04014-9. Online ahead of print.

ABSTRACT

BACKGROUND: This study aimed to develop and validate a hybrid deep learning (DL) model that integrates convolutional neural network (CNN) and vision transformer (ViT) architectures to predict distant metastasis (DM) in patients with non-small cell lung cancer (NSCLC) using 18F-FDG PET/CT images.

METHODS: A retrospective analysis was conducted on a cohort of consecutively registered patients who were newly diagnosed and untreated for NSCLC. A total of 167 patients with available PET/CT images were included in the analysis. DL features were extracted using a combination of CNN and ViT architectures, followed by feature selection, model construction, and evaluation of model performance using the receiver operating characteristic (ROC) and the area under the curve (AUC).

RESULTS: The ViT-based DL model exhibited strong predictive capabilities in both the training and validation cohorts, achieving AUCs of 0.824 and 0.830 for CT features, and 0.602 and 0.694 for PET features, respectively. Notably, the model that integrated both PET and CT features demonstrated a notable AUC of 0.882 in the validation cohort, outperforming models that utilized either PET or CT features alone. Furthermore, this model outperformed the CNN model (ResNet 50), which achieved an AUC of 0.752 [95% CI 0.613, 0.890], p < 0.05. Decision curve analysis further supported the efficacy of the ViT-based DL model.

CONCLUSION: The ViT-based DL developed in this study demonstrates considerable potential in predicting DM in patients with NSCLC, potentially informing the creation of personalized treatment strategies. Future validation through prospective studies with larger cohorts is necessary.

PMID:40779149 | DOI:10.1007/s12094-025-04014-9

Radiol Phys Technol. 2025 Aug 8. doi: 10.1007/s12194-025-00938-7. Online ahead of print.

ABSTRACT

Clinical magnetic resonance imaging (MRI) is a high-resolution tool widely used for detailed anatomical imaging. However, prolonged scan times often lead to motion artefacts and patient discomfort. Fast acquisition techniques can reduce scan times but often produce noisy, low-contrast images, compromising segmentation accuracy essential for diagnosis and treatment planning. To address these limitations, we developed an end-to-end framework that incorporates BIDS-based data organiser and anonymizer, a GAN-based MR image enhancement model (GAN-MRI), AssemblyNet for brain region segmentation, and an attention-residual U-Net with Guided loss for abdominal and thigh segmentation. Thirty brain scans (5,400 slices) and 32 abdominal (1,920 slices) and 55 thigh scans (2,200 slices) acquired from multiple MRI scanners (GE, Siemens, Toshiba) underwent evaluation. Image quality improved significantly, with SNR and CNR for brain scans increasing from 28.44 to 42.92 (p < 0.001) and 11.88 to 18.03 (p < 0.001), respectively. Abdominal scans exhibited SNR increases from 35.30 to 50.24 (p < 0.001) and CNR from 10,290.93 to 93,767.22 (p < 0.001). Double-blind evaluations highlighted improved visualisations of anatomical structures and bias field correction. Segmentation performance improved substantially in the thigh (muscle: + 21%, IMAT: + 9%) and abdominal regions (SSAT: + 1%, DSAT: + 2%, VAT: + 12%), while brain segmentation metrics remained largely stable, reflecting the robustness of the baseline model. Proposed framework is designed to handle data from multiple anatomies with variations from different MRI scanners and centres by enhancing MRI scan and improving segmentation accuracy, diagnostic precision and treatment planning while reducing scan times and maintaining patient comfort.

PMID:40779148 | DOI:10.1007/s12194-025-00938-7

Methods Mol Biol. 2025;2932:291-301. doi: 10.1007/978-1-0716-4566-6_16.

ABSTRACT

Somatic variant detection is an important step in the analysis of cancer genomes for basic research as well as precision oncology. Here, we review existing computational methods for identifying somatic mutations from tissue as well as liquid biopsy samples. We then describe steps to run VarNet (Krishnamachari et al., Nat Commun 13:4248, 2022), a variant caller using deep learning, to accurately identify single nucleotide variants (SNVs) and short insertion-deletion (indels) mutations from next-generation sequencing (NGS) of tumor tissue samples.

PMID:40779117 | DOI:10.1007/978-1-0716-4566-6_16

Methods Mol Biol. 2025;2932:273-289. doi: 10.1007/978-1-0716-4566-6_15.

ABSTRACT

Cancer precision medicine aims to identify the best course of treatment for an individual. To achieve this goal, two important questions include predicting the response of an individual to a treatment strategy and identifying molecular markers that determine the response. The rapid growth of large publicly available databases containing clinical and molecular characteristics of cancer-derived samples paired with their response to single or multiple drugs, has enabled the development of computational models to answer these questions. In recent years, various deep learning models have been proposed to predict the response to polytherapy and monotherapies. However, selecting among all available options or developing new models for a particular study requires careful considerations and best practices to avoid various pitfalls. In this chapter, and drawing from our own studies, we will discuss various important points for choosing, utilizing, and developing such deep learning tools.

PMID:40779116 | DOI:10.1007/978-1-0716-4566-6_15

Methods Mol Biol. 2025;2932:259-271. doi: 10.1007/978-1-0716-4566-6_14.

ABSTRACT

Over the past decade, convolutional neural networks (CNNs) have revolutionized predictive modeling of data containing spatial correlations, specifically excelling at image analysis tasks due to their embedded feature extraction and improved generalization. However, outside of image or sequence data, datasets typically lack the structural correlation needed to exploit the benefits of CNN modeling. This is especially true regarding anticancer drug sensitivity prediction tasks, as the data used is often tabular without any embedded information in the ordering or locations of the features when utilizing data other than DNA or RNA sequences. This chapter provides a computational procedure, REpresentation of Features as Images with NEighborhood Dependencies (REFINED), that maps high-dimensional feature vectors into compact 2D images suitable for CNN-based deep learning. The pairing of REFINED mappings with CNNs enables enhanced predictive performance through reduced model parameterization and improved embedded feature extraction as compared to fully connected alternatives utilizing the high-dimensional feature vectors.

PMID:40779115 | DOI:10.1007/978-1-0716-4566-6_14

J Phys Chem Lett. 2025 Aug 8:8418-8426. doi: 10.1021/acs.jpclett.5c01753. Online ahead of print.

ABSTRACT

Single-molecule detection of post-translational modifications (PTMs) such as phosphorylation plays a crucial role in early diagnosis of diseases and therapeutics development. Although single-molecule surface-enhanced Raman spectroscopy (SM-SERS) detection of PTMs has been demonstrated, the data analysis and detection accurracies were hindered by interference from citrate signals and lack of reference databases. Previous reports required complete coverage of the nanoparticle surface by analyte molecules to replace citrates, hampering the detection limit. Here, we developed a high-accuracy SM-SERS approach by combining a plasmonic particle-in-pore sensor to collect SM-SERS spectra of phosphorylation at Serine and Tyrosine, k-means-based clustering for citrate signal removal, and a one-dimensional convolutional neural network (1D-CNN) for phosphorylation identification. Significantly, we collected SM-SERS data with submonolayer analyte coverage of the particle surface and discriminated the phosphorylation in Serine and Tyrosine with over 95% and 97% accuracy, respectively. Finally, the 1D-CNN features were interpreted by a one-dimensional gradient feature weight and SM-SERS peak occurrence frequencies.

PMID:40778942 | DOI:10.1021/acs.jpclett.5c01753

JACC Cardiovasc Imaging. 2025 Aug 1:S1936-878X(25)00331-6. doi: 10.1016/j.jcmg.2025.06.010. Online ahead of print.

ABSTRACT

BACKGROUND: Single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) uses computed tomography (CT)-based attenuation correction (AC) to improve diagnostic accuracy. Deep learning (DL) has the potential to generate synthetic AC images, as an alternative to CT-based AC.

OBJECTIVES: This study evaluated whether DL-generated synthetic SPECT images could enhance accuracy of conventional SPECT MPI.

METHODS: Study investigators developed a DL model in a multicenter cohort of 4,894 patients from 4 sites to generate simulated SPECT AC images (DeepAC). The model was externally validated in 746 patients from 72 sites in a clinical trial (A Phase 3 Multicenter Study to Assess PET Imaging of Flurpiridaz F 18 Injection in Patients With CAD; NCT01347710) and in 320 patients from another external site. In the first external cohort, the study assessed the diagnostic accuracy for obstructive coronary artery disease (CAD)-defined as left main coronary artery stenosis ≥50% or ≥70% in other vessels-for total perfusion deficit (TPD). In the latter, the study completed change analysis and compared quantitative scores for AC, DeepAC, and nonattenuation correction (NC) with clinical scores.

RESULTS: In the first external cohort (mean age, 63 ± 9.5 years; 69.0% male), 206 patients (27.6%) had obstructive CAD. The area under the receiver-operating characteristic curve (AUC) of DeepAC TPD (0.77; 95% CI: 0.73-0.81) was higher than the NC TPD (AUC: 0.73; 95% CI: 0.69-0.77; P < 0.001). In the second external cohort, DeepAC quantitative scores had closer agreement with actual AC scores compared with NC.

CONCLUSIONS: In a multicenter external cohort, DeepAC improved prediction performance for obstructive CAD. This approach could enhance diagnostic accuracy in facilities using conventional SPECT systems without requiring additional equipment, imaging time, or radiation exposure.

PMID:40778900 | DOI:10.1016/j.jcmg.2025.06.010

JASA Express Lett. 2025 Aug 1;5(8):086001. doi: 10.1121/10.0038764.

ABSTRACT

Model-based deep learning approaches provide an alternative scheme to address the problem of the shortage of training data. However, performance degradation caused by sound speed profile (SSP) mismatch remains a critical challenge, particularly in shallow-water environments influenced by internal waves. In this paper, a simple range-dependent SSP model is integrated into the deep learning approach for source localization. The network trained on simulated data generated with the range-dependent SSP model performs well on validation data and generalizes to experimental test data after transfer learning with limited experimental samples.

PMID:40778845 | DOI:10.1121/10.0038764

Liver Int. 2025 Sep;45(9):e70274. doi: 10.1111/liv.70274.

ABSTRACT

BACKGROUND AND AIMS: The first variceal haemorrhage (FVH) is a life-threatening complication of liver cirrhosis that requires timely intervention; however, noninvasive tools for accurately predicting FVH remain limited. This study aimed to develop noninvasive, deep learning-enhanced computed tomographic angiography (CTA) models for early and accurate FVH prediction.

METHODS: This multi-centre retrospective study included 184 cirrhotic patients (FVH: n = 107, non-FVH: n = 77) enrolled from December 2014 to May 2022. Patients were randomly divided (7:3) into training and validation cohorts. CTA and clinical data were collected and analysed. A novel Vision-Transformer (ViT) network, combined with reinforcement learning (RL), was applied to CTA images to predict FVH and was compared with convolutional neural networks (CNNs). Models were evaluated using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA), and feature importance was determined from model coefficients and gradients.

RESULTS: The ViT + RL* model demonstrated superior diagnostic performance, achieving an AUC of 0.985 (95% CI, 0.955-1.0) in the validation cohort and 0.956 (95% CI, 0.919-0.988) in the training cohort, outperforming traditional CNNs. DCA and the area under the curve confirmed the enhanced clinical utility of the ViT + RL* model compared to CNNs; the ViT + RL* model highlighted critical regions in the liver, spleen, oesophageal lumen, and abdominal vessels. Meanwhile, clinical data identified creatinine and prothrombin time as potential predictive factors, with moderate predictive performance.

CONCLUSIONS: The novel deep learning-enhanced CTA models offer a robust, non-invasive method for predicting FVH, with the ViT + RL* model demonstrating excellent efficacy; thus providing a valuable tool for early risk stratification in cirrhotic patients.

PMID:40778828 | DOI:10.1111/liv.70274

Chemistry. 2025 Aug 8:e02067. doi: 10.1002/chem.202502067. Online ahead of print.

ABSTRACT

The inherently low sensitivity of solution-state Nuclear Magnetic Resonance (NMR) has long limited its ability to characterize transient biomolecular states at atomic resolution. While dissolution dynamic nuclear polarization (dDNP) offers a compensating signal enhancement, its broader use has been hindered by rapid polarization decay, causing severe spectral distortion. Here, we introduce HyperW-Decon, an approach that enables high-sensitivity, high-resolution NMR of biomolecules in solution. HyperW-Decon combines two key aspects: (i) the use of hyperpolarized water (HyperW) to transfer polarization to proteins through rapid proton exchange; and (ii) a theory-driven, machine learning (ML)-based deconvolution method that corrects polarization-induced artifacts without requiring external reference signals. This approach is based on a first-principles understanding of dDNP line shapes and delivers a scalable solution to spectral distortion. Applied to intrinsically disordered proteins (IDPs) involved in biomineralization, HyperW-Decon reveals previously inaccessible, short-lived ion-peptide encounter complexes with residue resolution.

PMID:40778633 | DOI:10.1002/chem.202502067

Data Brief. 2025 Jul 16;62:111890. doi: 10.1016/j.dib.2025.111890. eCollection 2025 Oct.

ABSTRACT

This dataset presents a comprehensive multimodal collection of data acquired from the chewing of beef and plant-based burger patties using both human participants (Invivo) and a biomimicking robotic chewing device (Invitro). The primary objective of the data collection was to discover relationships regarding the change in food bolus properties with the number of robotic chewing cycles as the human swallowing threshold is achieved, which will facilitate the development of deep learning models capable of predicting mechanical and textural properties of chewed food boluses from images. In the in-vivo experiments, expectorated bolus samples were collected from three healthy adult male participants, who chewed food samples until just before swallowing. The chewed boluses were then imaged using a 12MP camera and a flatbed scanner, followed by Texture Profile Analysis (TPA) to measure texture parameters. The dataset comprises two main folders Invivo and Invitro. The Invivo data thus comprises high-resolution images and corresponding TPA metrics at the near-swallowing stage. In the Invitro experiments, a 3 degree of freedom linkage chewing robot (ChewBot) with a soft robotic oral cavity was used to simulate human mastication. The robot performed controlled mastication using different molar trajectories that varied in lateral shearing effect. Food samples were chewed for up to 40 chewing cycles, with artificial saliva introduced at 10 % of the food's weight. For each experimental condition, the dataset includes real-time images captured immediately after each the robotic chewing cycle, force profiles recorded at 100 ms intervals during the chewing, and TPA metrics of the resulting bolus after every 5 chewing cycles. This dataset has significant reuse potential in various fields. In food science, it can support studies on the mechanical breakdown of meat and meat alternatives, aiding in the reformulation of plant-based foods to better mimic desirable animal-based food textures. This dataset supports rehabilitation in health sciences by aiding personalized diet design for individuals with jaw disorders or dysphagia and guiding texture-appropriate menu options for patients. In robotics and artificial mastication, it informs the development of chewing systems. It also enables machine learning applications for predicting food texture from images, allowing automated, non-invasive analysis.

PMID:40778379 | PMC:PMC12329220 | DOI:10.1016/j.dib.2025.111890

Ophthalmol Sci. 2025 Jun 19;5(6):100861. doi: 10.1016/j.xops.2025.100861. eCollection 2025 Nov-Dec.

ABSTRACT

CLINICAL RELEVANCE: Corneal ulcers cause preventable blindness in >2 million individuals annually, primarily affecting low- and middle-income countries. Prompt and accurate pathogen identification is essential for targeted antimicrobial treatment, yet current diagnostic methods are costly and slow and require specialized expertise, limiting accessibility.

METHODS: We systematically reviewed literature published from 2017 to 2024, identifying 37 studies that developed or validated artificial intelligence (AI) models for pathogen detection and related classification tasks in infectious keratitis. The studies were analyzed for model types, input modalities, datasets, ground truth determination methods, and validation practices.

RESULTS: Artificial intelligence models demonstrated promising accuracy in pathogen detection using image interpretation techniques. Common limitations included limited generalizability, lack of diverse datasets, absence of multilabeled classification methods, and variability in ground truth standards. Most studies relied on single-center retrospective datasets, limiting applicability in routine clinical practice.

CONCLUSIONS: Artificial intelligence shows significant potential to improve pathogen detection in infectious keratitis, enhancing both diagnostic accuracy and accessibility globally. Future research should address identified limitations by increasing dataset diversity, adopting multilabel classification, implementing prospective and multicenter validations, and standardizing ground truth definitions.

FINANCIAL DISCLOSURES: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

PMID:40778364 | PMC:PMC12329105 | DOI:10.1016/j.xops.2025.100861

Ophthalmol Sci. 2025 Jun 16;5(6):100852. doi: 10.1016/j.xops.2025.100852. eCollection 2025 Nov-Dec.

ABSTRACT

PURPOSE: We present an algorithm to segment subretinal fluid (SRF) on individual B-scan slices in patients with rhegmatogenous retinal detachment (RRD). Particular attention is paid to robustness, with a fivefold cross-validation approach and a hold-out test set.

DESIGN: Retrospective, cross-sectional study.

PARTICIPANTS: A total of 3819 B-scan slices across 98 time points from 45 patients were used in this study.

METHODS: Subretinal fluid was segmented on all scans. A base SegFormer model, pretrained on 4 massive data sets, was further trained on raw B-scans from the retinal OCT fluid challenge data set of 4532 slices: an open data set of intraretinal fluid, SRF, and pigment epithelium detachment. When adequate performance was reached, transfer learning was used to train the model on our in-house data set, to segment SRF by generating a pixel-wise mask of presence/absence of SRF. A fivefold cross-validation approach was used, with an additional hold-out test set. All folds were first trained and cross-validated and then additionally tested on the hold-out set. Mean (averaged across images) and total (summed across all pixels, irrespective of image) Dice coefficients were calculated for each fold.

MAIN OUTCOME MEASURES: Subretinal fluid volume after surgical intervention for RRD.

RESULTS: The average total Dice coefficient across the validation folds was 0.92, the average mean Dice coefficient was 0.82, and the median Dice was 0.92. For the test set, the average total Dice coefficient was 0.94, the average mean Dice coefficient was 0.82, and the median Dice was 0.92. The model showed strong interfold consistency on the hold-out set, with a standard deviation of only 0.03.

CONCLUSIONS: The SegFormer model for SRF segmentation demonstrates a strong ability to segment SRF. This result holds up to cross-validation and hold-out testing, across all folds. The model is available open-source online.

FINANCIAL DISCLOSURES: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

PMID:40778358 | PMC:PMC12329092 | DOI:10.1016/j.xops.2025.100852

Front Endocrinol (Lausanne). 2025 Jul 24;16:1634875. doi: 10.3389/fendo.2025.1634875. eCollection 2025.

ABSTRACT

OBJECTIVE: This study aimed to evaluate the value of constructing a multimodal deep-learning video model based on 2D ultrasound and contrast-enhanced ultrasound (CEUS) dynamic video for the preoperative prediction of OLNM in papillary thyroid carcinoma (PTC) patients.

METHODS: A retrospective analysis was conducted on 396 cases of clinically lymph node-negative PTC cases with ultrasound images collected between January and September 2023. Five representative deep learning architectures were pre-trained to construct deep learning static image models (DL_image), CEUS dynamic video models (DL_CEUSvideo), and combined models (DL_combined). The area under the receiver operating characteristic curve (AUC) was used to evaluate model performance, with comparisons made using the Delong test. A P-value of less than 0.05 was considered statistically significant.

RESULTS: The DL_CEUSvideo, DL_image, and DL_combined models were successfully developed and demonstrated. The AUC values were 0.826 (95% CI: 0.771-0.881), 0.759 (95% CI: 0.690-0.828), and 0.926 (95% CI: 0.891-0.962) in the training set, and 0.701 (95% CI: 0.589-0.813), 0.624 (95% CI: 0.502-0.745), and 0.734 (95% CI: 0.627-0.842) in the test set. Finally, sensitivity, specificity, and accuracy for the DL_CEUSvideo, DL_image, and DL_combined models were 0.836, 0.671, 0.704; 0.673, 0.716, 0.707; and 0.818, 0.902, 0.886 in the training set, and 0.556, 0.775, 0.724; 0.556, 0.674, 0.647; and 0.704, 0.663, 0.672 in the test set, respectively.

CONCLUSION: These results demonstrated that the multimodal deep learning dynamic video model could preoperatively predict OLNM in PTC patients. The DL_CEUSvideo model outperformed the DL_image model, while the DL_combined model significantly enhanced sensitivity without compromising specificity.

PMID:40778281 | PMC:PMC12329689 | DOI:10.3389/fendo.2025.1634875