Abdom Radiol (NY). 2025 Apr 1. doi: 10.1007/s00261-025-04886-z. Online ahead of print.

ABSTRACT

PURPOSE: To develop and validate a deep learning-based feature ensemble model using multiparametric magnetic resonance imaging (MRI) for predicting tumor budding (TB) grading in patients with rectal cancer (RC).

METHODS: A retrospective cohort of 458 patients with pathologically confirmed rectal cancer (RC) from three institutions was included. Among them, 355 patients from Center 1 were divided into two groups at a 7:3 ratio: the training cohort (n = 248) and the internal validation cohort (n = 107). An additional 103 patients from two other centers served as the external validation cohort. Deep learning models were constructed for T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) based on the CrossFormer architecture, and deep learning features were extracted. Subsequently, a feature ensemble module based on the attention mechanism of Transformer was used to capture spatial interactions between different imaging sequences, creating a multiparametric ensemble model. The predictive performance of each model was evaluated using the area under the curve (AUC), calibration curves, and decision curve analysis (DCA).

RESULTS: The deep learning model based on T2WI achieved AUC values of 0.789 (95% CI: 0.680-0.900) and 0.720 (95% CI: 0.591-0.849) in the internal and external validation cohorts, respectively. The deep learning model based on DWI had AUC values of 0.806 (95% CI: 0.705-0.908) and 0.772 (95% CI: 0.657-0.887) in the internal and external validation cohorts, respectively. The multiparametric ensemble model demonstrated superior performance, with AUC values of 0.868 (95% CI: 0.775-0.960) in the internal validation cohort and 0.839 (95% CI: 0.743-0.935) in the external validation cohort. DeLong test showed that the differences in AUC values among the models were not statistically significant in both the internal and external test sets (P > 0.05). The DCA curve demonstrated that within the 10-80% threshold range, the fusion model provided significantly higher clinical net benefit compared to other models.

CONCLUSION: Compared to single-sequence deep learning models, the attention mechanism-based multiparametric MRI fusion model enables more effective individualized prediction of TB grading in RC patients. It offers valuable guidance for treatment selection and prognostic evaluation while providing imaging-based support for personalized postoperative follow-up adjustments.

PMID:40167646 | DOI:10.1007/s00261-025-04886-z

Abdom Radiol (NY). 2025 Apr 1. doi: 10.1007/s00261-025-04887-y. Online ahead of print.

ABSTRACT

OBJECTIVES: To train and validate segmentation models for automated segmentation of gallbladder cancer (GBC) lesions from contrast-enhanced CT images.

MATERIALS AND METHODS: This retrospective study comprised consecutive patients with pathologically proven treatment naïve GBC who underwent a contrast-enhanced CT scan at four different tertiary care referral hospitals. The training and validation cohort comprised CT scans of 317 patients (center 1). The internal test cohort comprised a temporally independent cohort (n = 29) from center 1 (internal test 1). The external test cohort comprised CT scans from three centers [ (n = 85)]. We trained the state-of-the-art 2D and 3D image segmentation models, SAM Adapter, MedSAM, 3D TransUNet, SAM-Med3D, and 3D-nnU-Net, for automated segmentation of the GBC. The models' performance for GBC segmentation on the test datasets was assessed via dice score and intersection over union (IoU) using manual segmentation as the reference standard.

RESULTS: The 2D models performed better than 3D models. Overall, MedSAM achieved the highest dice and IoU scores on both the internal [mean dice (SD) 0.776 (0.106) and mean IoU 0.653 (0.133)] and external [mean dice (SD) 0.763 (0.098) and mean IoU 0.637 (0.116)] test sets. Among the 3D models, TransUNet showed the best segmentation performance with mean dice (SD) and IoU (SD) of 0.479 (0.268) and 0.356 (0.235) in the internal test and 0.409 (0.339) and 0.317 (0.283) in the external test sets. The segmentation performance was not associated with GBC morphology. There was weak correlation between the dice/IoU and the size of the GBC lesions for any segmentation model.

CONCLUSION: We trained 2D and 3D GBC segmentation models on a large dataset and validated these models on external datasets. MedSAM, a 2D prompt-based foundational model, achieved the best segmentation performance.

PMID:40167645 | DOI:10.1007/s00261-025-04887-y

Abdom Radiol (NY). 2025 Apr 1. doi: 10.1007/s00261-025-04897-w. Online ahead of print.

ABSTRACT

Artificial intelligence (AI) may have the potential to improve existing diagnostic challenges in endometriosis imaging. To better direct future research, this descriptive review summarizes the general landscape of AI applications in endometriosis imaging. Articles from PubMed were selected to represent different approaches to AI applications in endometriosis imaging. Current endometriosis imaging literature focuses on AI applications in ultrasound (US) and magnetic resonance imaging (MRI). Most studies use US data, with MRI studies being limited at present. The majority of US studies employ transvaginal ultrasound (TVUS) data and aim to detect deep endometriosis implants, adenomyosis, endometriomas, and secondary signs of endometriosis. Most MRI studies evaluate endometriosis disease diagnosis and segmentation. Some studies analyze multi-modal methods for endometriosis imaging, combining US and MRI data or using imaging data in combination with clinical data. Current literature lacks generalizability and standardization. Most studies in this review utilize small sample sizes with retrospective approaches and single-center data. Existing models only focus on narrow disease detection or diagnosis questions and lack standardized ground truth. Overall, AI applications in endometriosis imaging analysis are in their early stages, and continued research is essential to develop and enhance these models.

PMID:40167644 | DOI:10.1007/s00261-025-04897-w

JASA Express Lett. 2025 Apr 1;5(4):040801. doi: 10.1121/10.0036312.

ABSTRACT

A generative deep learning framework is introduced for ray-based ocean acoustic tomography (OAT), an inverse problem for estimating sound speed profiles (SSP) based on arrival-times measurements between multiple acoustic transducers, which is typically ill-posed. This framework relies on a robust low-dimensional parametrization of the expected SSP variations using a variational autoencoder and a linear dynamical model as further regularization. This framework was tested using SSP variations simulated by a regional ocean model with submesoscale permitting horizontal resolution and various transducer configurations spanning the upper ocean over short propagation ranges and was found to outperform conventional linear least squares formulations of OAT.

PMID:40167492 | DOI:10.1121/10.0036312

J Synchrotron Radiat. 2025 May 1. doi: 10.1107/S1600577525001833. Online ahead of print.

ABSTRACT

In bone-imaging research, in situ synchrotron radiation micro-computed tomography (SRµCT) mechanical tests are used to investigate the mechanical properties of bone in relation to its microstructure. Low-dose computed tomography (CT) is used to preserve bone's mechanical properties from radiation damage, though it increases noise. To reduce this noise, the self-supervised deep learning method Noise2Inverse was used on low-dose SRµCT images where segmentation using traditional thresholding techniques was not possible. Simulated-dose datasets were created by sampling projection data at full, one-half, one-third, one-fourth and one-sixth frequencies of an in situ SRµCT mechanical test. After convolutional neural networks were trained, Noise2Inverse performance on all dose simulations was assessed visually and by analyzing bone microstructural features. Visually, high image quality was recovered for each simulated dose. Lacunae volume, lacunae aspect ratio and mineralization distributions shifted slightly in full, one-half and one-third dose network results, but were distorted in one-fourth and one-sixth dose network results. Following this, new models were trained using a larger dataset to determine differences between full dose and one-third dose simulations. Significant changes were found for all parameters of bone microstructure, indicating that a separate validation scan may be necessary to apply this technique for microstructure quantification. Noise present during data acquisition from the testing setup was determined to be the primary source of concern for Noise2Inverse viability. While these limitations exist, incorporating dose calculations and optimal imaging parameters enables self-supervised deep learning methods such as Noise2Inverse to be integrated into existing experiments to decrease radiation dose.

PMID:40167487 | DOI:10.1107/S1600577525001833

ArXiv [Preprint]. 2025 Mar 20:arXiv:2503.16706v1.

ABSTRACT

A recently proposed scatter-window and deep learning-based attenuation compensation (AC) method for myocardial perfusion imaging (MPI) by single-photon emission computed tomography (SPECT), namely CTLESS, demonstrated promising performance on the clinical task of myocardial perfusion defect detection with retrospective data acquired on SPECT scanners from a single vendor. For clinical translation of CTLESS, it is important to assess the generalizability of CTLESS across different SPECT scanners. For this purpose, we conducted a virtual imaging trial, titled in silico imaging trial to assess generalizability (ISIT-GEN). ISIT-GEN assessed the generalizability of CTLESS on the cardiac perfusion defect detection task across SPECT scanners from three different vendors. The performance of CTLESS was compared with a standard-of-care CT-based AC (CTAC) method and a no-attenuation compensation (NAC) method using an anthropomorphic model observer. We observed that CTLESS had receiver operating characteristic (ROC) curves and area under the ROC curves similar to those of CTAC. Further, CTLESS was observed to significantly outperform the NAC method across three scanners. These results are suggestive of the inter-scanner generalizability of CTLESS and motivate further clinical evaluations. The study also highlights the value of using in silico imaging trials to assess the generalizability of deep learning-based AC methods feasibly and rigorously.

PMID:40166744 | PMC:PMC11957238

medRxiv [Preprint]. 2025 Mar 20:2025.03.18.25324199. doi: 10.1101/2025.03.18.25324199.

ABSTRACT

BACKGROUND: Accurate delineation of orodental structures on radiotherapy CT images is essential for dosimetric assessments and dental decisions. We propose a deep-learning auto-segmentation framework for individual teeth and mandible/maxilla sub-volumes aligned with the ClinRad ORN staging system.

METHODS: Mandible and maxilla sub-volumes were manually defined, differentiating between alveolar and basal regions, and teeth were labelled individually. For each task, a DL segmentation model was independently trained. A Swin UNETR-based model was used for the mandible sub-volumes. For the smaller structures (e.g., teeth and maxilla sub-volumes) a two-stage segmentation model first used the ResUNet to segment the entire teeth and maxilla regions as a single ROI that was then used to crop the image input of the Swin UNETR. In addition to segmentation accuracy and geometric precision, a dosimetric comparison was made between manual and model-predicted segmentations.

RESULTS: Segmentation performance varied across sub-volumes - mean Dice values of 0.85 (mandible basal), 0.82 (mandible alveolar), 0.78 (maxilla alveolar), 0.80 (upper central teeth), 0.69 (upper premolars), 0.76 (upper molars), 0.76 (lower central teeth), 0.70 (lower premolars), 0.71 (lower molars) - and exhibited limited applicability in segmenting teeth and sub-volumes often absent in the data. Only the maxilla alveolar central sub-volume showed a statistically significant dosimetric difference (Bonferroni-adjusted p-value = 0.02).

CONCLUSION: We present a novel DL-based auto-segmentation framework of orodental structures, enabling spatial localization of dose-related differences in the jaw. This tool enhances image-based bone injury detection, including ORN, and improves clinical decision-making in radiation oncology and dental care for head and neck cancer patients.

PMID:40166584 | PMC:PMC11957087 | DOI:10.1101/2025.03.18.25324199

medRxiv [Preprint]. 2025 Mar 20:2025.03.19.25324270. doi: 10.1101/2025.03.19.25324270.

ABSTRACT

Antimicrobial resistance poses a major global threat due to the diminishing efficacy of current treatments and limited new therapies. Combination therapy with existing drugs offers a promising solution, yet current empirical methods often lead to suboptimal efficacy and inadvertent toxicity. The high cost of experimentally testing numerous combinations underscores the need for data-driven methods to streamline treatment design. We introduce CALMA, an approach that predicts the potency and toxicity of multi-drug combinations in Escherichia coli and Mycobacterium tuberculosis . CALMA identified synergistic antimicrobial combinations involving vancomycin and isoniazid that were antagonistic for toxicity, which were validated using in vitro cell viability assays in human cell lines and through mining of patient health records that showed reduced side effects in patients taking combinations identified by CALMA. By combining mechanistic modelling with deep learning, CALMA improves the interpretability of neural networks, identifies key pathways influencing drug interactions, and prioritizes combinations with enhanced potency and reduced toxicity.

PMID:40166569 | PMC:PMC11957163 | DOI:10.1101/2025.03.19.25324270

medRxiv [Preprint]. 2025 Mar 19:2025.03.18.25324215. doi: 10.1101/2025.03.18.25324215.

ABSTRACT

BACKGROUND: Accurate measurement of echocardiographic parameters is crucial for the diagnosis of cardiovascular disease and tracking of change over time, however manual assessment is time-consuming and can be imprecise. Artificial intelligence (AI) has the potential to reduce clinician burden by automating the time-intensive task of comprehensive measurement of echocardiographic parameters.

METHODS: We developed and validated open-sourced deep learning semantic segmentation models for the automated measurement of 18 anatomic and Doppler measurements in echocardiography. The outputs of segmentation models were compared to sonographer measurements from two institutions to access accuracy and precision.

RESULTS: We utilized 877,983 echocardiographic measurements from 155,215 studies from Cedars-Sinai Medical Center (CSMC) to develop EchoNet-Measurements, an open-source deep learning model for echocardiographic annotation. The models demonstrated a good correlation when compared with sonographer measurements from held-out data from CSMC and an independent external validation dataset from Stanford Healthcare (SHC). Measurements across all nine B-mode and nine Doppler measurements had high accuracy (an overall R 2 of 0.967 (0.965 - 0.970) in the held-out CSMC dataset and 0.987 (0.984 - 0.989) in the SHC dataset). When evaluated end-to-end on a temporally distinct 2,103 studies at CSMC, EchoNet-Measurements performed well an overall R2 of 0.981 (0.976 - 0.984). Performance was consistent across patient characteristics including sex and atrial fibrillation status.

CONCLUSION: EchoNet-Measurement achieves high accuracy in automated echocardiographic measurement that is comparable to expert sonographers. This open-source model provides the foundation for future developments in AI applied to echocardiography.

CLINICAL PERSPECTIVE: What Is New?: We developed EchoNet-Measurements, the first publicly available deep learning framework for comprehensive automated echocardiographic measurements.We assessed the performance of EchoNet-Measurements, showing good precision and accuracy compared to human sonographers and cardiologists across multiple healthcare systems.What Are the Clinical Implications?: Deep-learning automated echocardiographic measurements can be conducted in a fraction of a second, reducing the time burden on sonographers and standardizing measurements, and potentially enhance reproducibility and diagnostic reliability.This open-source model provides broad opportunities for widespread adoption in both clinical use and research, including in resource-limited settings.

PMID:40166567 | PMC:PMC11957091 | DOI:10.1101/2025.03.18.25324215

medRxiv [Preprint]. 2025 Mar 20:2025.03.19.25324269. doi: 10.1101/2025.03.19.25324269.

ABSTRACT

Perivascular spaces (PVS) are cerebrospinal fluid-filled tunnels around brain blood vessels, crucial for the functions of the glymphatic system. Changes in PVS have been linked to vascular diseases and aging, necessitating accurate segmentation for further study. PVS segmentation poses challenges due to their small size, varying MRI appearances, and the scarcity of annotated data. We present a finely segmented PVS dataset from T2-weighted MRI scans, sourced from the Human Connectome Project Aging (HCP-Aging), encompassing 200 subjects aged 30 to 100. Our approach utilizes a combination of unsupervised and deep learning techniques with manual corrections to ensure high accuracy. This dataset aims to facilitate research on PVS dynamics across different ages and to explore their link to cognitive decline. It also supports the development of advanced image segmentation algorithms, contributing to improved medical imaging automation and the early detection of neurodegenerative diseases.

PMID:40166557 | PMC:PMC11957161 | DOI:10.1101/2025.03.19.25324269

medRxiv [Preprint]. 2025 Mar 19:2025.03.18.25323918. doi: 10.1101/2025.03.18.25323918.

ABSTRACT

BACKGROUND: Accurate evaluation of aortic regurgitation (AR) severity is necessary for early detection and chronic disease management. AR is most commonly assessed by Doppler echocardiography, however limitations remain given variable image quality and need to integrate information from multiple views. This study developed and validated a deep learning model for automated AR severity assessment from multi-view color Doppler videos.

METHODS: We developed a video-based convolutional neural network (R2+1D) to classify AR severity using color Doppler echocardiography videos from five standard views: parasternal long-axis (PLAX), PLAX-aortic valve focus, apical three-chamber (A3C), A3C-aortic valve focus, and apical five-chamber (A5C). The model was trained on 47,638 videos from 32,396 studies (23,240 unique patients) from Cedars-Sinai Medical Center (CSMC) and externally validated on 3369 videos from 1504 studies (1493 unique patients) from Stanford Healthcare Center (SHC).

RESULTS: Combining assessments from multiple views, the EchoNet-AR model achieved excellent identification of both at least moderate AR (AUC 0.95, [95% CI 0.94-0.96]) and severe AR (AUC 0.97, [95% CI 0.96 - 0.98]). This performance was consistent in the external SHC validation cohort for both at least moderate AR (AUC 0.92, [95% CI 0.88-0.96]) and severe AR (AUC 0.94, [95% CI 0.89-0.98]). Subgroup analysis showed robust model performance across varying image quality, valve morphologies, and patient demographics. Saliency map visualizations demonstrated that the model focused on the proximal flow convergence zone and vena contracta, appropriately narrowing on hemodynamically significant regions.

CONCLUSION: The EchoNet-AR model accurately classifies AR severity and synthesizes information across multiple echocardiographic views with robust generalizability in an external cohort. The model shows potential as an automated clinical decision support tool for AR assessment, however clinical interpretation remains essential, particularly in complex cases with multiple valve pathologies or altered hemodynamics.

PMID:40166551 | PMC:PMC11957077 | DOI:10.1101/2025.03.18.25323918

Front Artif Intell. 2025 Mar 17;8:1496109. doi: 10.3389/frai.2025.1496109. eCollection 2025.

ABSTRACT

BACKGROUND: The diagnostic power of exercise stress electrocardiography (ExECG) remains limited. We aimed to construct an artificial intelligence (AI)-based method to enhance ExECG performance to identify patients with significant coronary artery disease (CAD).

METHODS: We retrospectively collected 818 patients who underwent both ExECG and coronary angiography (CAG) within 6 months. The mean age was 57.0 ± 10.1 years, and 614 (75%) were male patients. Significant coronary artery disease was seen in 369 (43.8%) CAG reports. We also included 197 individuals with normal ExECG and low risk of CAD. A convolutional recurrent neural network algorithm, integrating electrocardiographic (ECG) signals and features from ExECG reports, was developed to predict the risk of significant CAD. We also investigated the optimal number of inputted ECG signal slices and features and the weighting of features for model performance.

RESULTS: Using the data of patients undergoing CAG for training and test sets, our algorithm had an area under the curve, sensitivity, and specificity of 0.74, 0.86, and 0.47, respectively, which increased to 0.83, 0.89, and 0.60, respectively, after enrolling 197 subjects with low risk of CAD. Three ECG signal slices and 12 features yielded optimal performance metrics. The principal predictive feature variables were sex, maximum heart rate, and ST/HR index. Our model generated results within one minute after completing ExECG.

CONCLUSION: The multimodal AI algorithm, leveraging deep learning techniques, efficiently and accurately identifies patients with significant CAD using ExECG data, aiding clinical screening in both symptomatic and asymptomatic patients. Nevertheless, the specificity remains moderate (0.60), suggesting a potential for false positives and highlighting the need for further investigation.

PMID:40166362 | PMC:PMC11955648 | DOI:10.3389/frai.2025.1496109

bioRxiv [Preprint]. 2025 Mar 17:2025.03.14.643307. doi: 10.1101/2025.03.14.643307.

ABSTRACT

Due to the favorable chemical properties of mirrored chiral centers (such as improved stability, bioavailability, and membrane permeability) the computational design of D-peptides targeting biological L-proteins is a valuable area of research. To design these structures in silico , a computational workflow should correctly dock and fold a peptide while maintaining chiral centers. The latest AlphaFold 3 (AF3) from Abramson et al. (2024) enforces a strict chiral violation penalty to maintain chiral centers from model inputs and is reported to have a low chiral violation rate of only 4.4% on a PoseBusters benchmark containing diverse chiral molecules. Herein, we report the results of 3,255 experiments with AF3 to evaluate its ability to predict the fold, chirality, and binding pose of D-peptides in heterochiral complexes. Despite our inputs specifying explicit D-stereocenters, we report that the AF3 chiral violation for D-peptide binders is much higher at 51% across all evaluated predictions; on average the model is as accurate as chance (random chirality choice, L or D, for each peptide residue). Increasing the number of seeds failed to improve this violation rate. The AF3 predictions exhibit incorrect folds and binding poses, with D-peptides commonly oriented incorrectly in the L-protein binding pocket. Confidence metrics returned by AF3 also fail to distinguish predictions with low chirality violation and correct docking vs. predictions with high chirality violation and incorrect docking. We conclude that AF3 is a poor predictor of D-peptide chirality, fold, and binding pose.

SUMMARY: A crucial task in computational protein design is predicting fold, as this property determines the structure and function of a protein. Abramson et al. 1 published in Nature on AlphaFold 3 (AF3), a powerful deep learning framework for predicting chemical structures in both bound and unbound states. This architecture is tuned to respect chiral centers, which are atoms (in proteins, backbone α -carbons) covalently bound to four different chemical species 2 . These centers adopt two non-superposable forms, often called "handedness," termed L (all biological proteins adopt this form) and D (the mirror image of L). L and D chiral centers exert significant influence on chemical function; changing the chirality of even a single residue can dramatically alter chemical properties such as enantioselective binding (e.g., antifolate resistance 3 ) and stability 4 . Additionally, D-peptides (small proteins containing exclusively D chiral centers) exhibit many advantages compared to their L-peptide counterparts, such as protease evasion 5 , and are therefore therapeutically relevant modalities. Due to vastly differing chemical properties, an algorithm should respect chiral center inputs and exhibit an error rate of 0%. Although Abramson et al. 1 reports a low 4.4% chirality violation across diverse chiral centers, we have found that the chiral violation rate for D-peptides with D chiral center inputs explicitly specified is much higher at 51%. Increasing the number of seeds fails to improve this rate. Our data highlights a crucial structural prediction error in AF3 and demonstrates the model is as accurate on average as chance (random chirality choice, L or D, for each peptide residue). Compared to empirical structures, AF3 is also highly inaccurate when folding and docking D-peptide:L-protein complexes. The failure of AF3 to accurately predict these chemical interactions indicates more work is need for high-quality prediction of D-peptides.

PMID:40166350 | PMC:PMC11956919 | DOI:10.1101/2025.03.14.643307

bioRxiv [Preprint]. 2025 Mar 21:2025.03.21.644611. doi: 10.1101/2025.03.21.644611.

ABSTRACT

Recording activity from large cell populations in deep neural circuits is essential for understanding brain function. Three-photon (3P) imaging is an emerging technology that allows for imaging of structure and function in subcortical brain structures. However, increased tissue heating, as well as the low repetition rate sources inherent to 3P imaging, have limited the fields of view (FOV) to areas of ≤0.3 mm 2 . Here we present a Large Imaging Field of view Three-photon (LIFT) microscope with a FOV of [gt]3 mm 2 . LIFT combines high numerical aperture (NA) optimized sampling, using a custom scanning module, with deep learning-based denoising, to enable population imaging in deep brain regions. We demonstrate non-invasive calcium imaging in the mouse brain from >1500 cells across CA1, the surrounding white matter, and adjacent deep layers of the cortex, and show population imaging with high signal-to-noise in the rat cortex at a depth of 1.2 mm. The LIFT microscope was built with all off-the-shelf components and allows for a flexible choice of imaging scale and rate.

PMID:40166349 | PMC:PMC11957121 | DOI:10.1101/2025.03.21.644611

bioRxiv [Preprint]. 2025 Mar 20:2025.01.14.633078. doi: 10.1101/2025.01.14.633078.

ABSTRACT

Post-translational modifications (PTMs) play a crucial role in allowing cells to expand the functionality of their proteins and adaptively regulate their signaling pathways. Defects in PTMs have been linked to numerous developmental disorders and human diseases, including cancer, diabetes, heart, neurodegenerative and metabolic diseases. PTMs are important targets in drug discovery, as they can significantly influence various aspects of drug interactions including binding affinity. The structural consequences of PTMs, such as phosphorylation-induced conformational changes or their effects on ligand binding affinity, have historically been challenging to study on a large scale, primarily due to reliance on experimental methods. Recent advancements in computational power and artificial intelligence, particularly in deep learning algorithms and protein structure prediction tools like AlphaFold3, have opened new possibilities for exploring the structural context of interactions between PTMs and drugs. These AI-driven methods enable accurate modeling of protein structures including prediction of PTM-modified regions and simulation of ligand-binding dynamics on a large scale. In this work, we identified small molecule binding-associated PTMs that can influence drug binding across all human proteins listed as small molecule targets in the DrugDomain database, which we developed recently. 6,131 identified PTMs were mapped to structural domains from Evolutionary Classification of Protein Domains (ECOD) database. Scientific contribution. Using recent AI-based approaches for protein structure prediction (AlphaFold3, RoseTTAFold All-Atom, Chai-1), we generated 14,178 models of PTM-modified human proteins with docked ligands. Our results demonstrate that these methods can predict PTM effects on small molecule binding, but precise evaluation of their accuracy requires a much larger benchmarking set. We also found that phosphorylation of NADPH-Cytochrome P450 Reductase, observed in cervical and lung cancer, causes significant structural disruption in the binding pocket, potentially impairing protein function. All data and generated models are available from DrugDomain database v1.1 ( http://prodata.swmed.edu/DrugDomain/ ) and GitHub ( https://github.com/kirmedvedev/DrugDomain ). This resource is the first to our knowledge in offering structural context for small molecule binding-associated PTMs on a large scale.

PMID:40166291 | PMC:PMC11956905 | DOI:10.1101/2025.01.14.633078

Sci Rep. 2025 Mar 31;15(1):10956. doi: 10.1038/s41598-024-81478-9.

ABSTRACT

To address the limitations of traditional partial discharge (PD) detection methods for switchgear, which fail to meet the requirements for real-time monitoring, rapid assessment, sample fusion, and joint analysis in practical applications, a joint PD recognition method of switchgear based on edge computing and deep learning is proposed. An edge collaborative defect identification architecture for switchgear is constructed, which includes the terminal device side, terminal collection side, edge-computing side, and cloud-computing side. The PD signal of switchgear is extracted based on UHF sensor and broadband pulse current sensor on the terminal collection side. Multidimensional features are obtained from these signals and a high-dimensional feature space is constructed based on feature extraction and dimensionality reduction on the edge-computing side. On the cloud side, the deep belief network (DBN)-based switchgear PD defect identification method is proposed and the PD samples acquired on the edge side are transmitted in real time to the cloud for training. Upon completion of the training, the resulting model is transmitted back to the edge side for inference, thereby facilitating real-time joint analysis of PD defects across multiple switchgear units. Verification of the proposed method is conducted using PD samples simulated in the laboratory. The results indicate that the DBN proposed in this paper can recognize PDs in switchgear with an accuracy of 88.03%, and under the edge computing architecture, the training time of the switchgear PD defect type classifier can be reduced by 44.28%, overcoming the challenges associated with traditional diagnostic models, which are characterized by long training durations, low identification efficiency, and weak collaborative analysis capabilities.

PMID:40164608 | DOI:10.1038/s41598-024-81478-9

Acad Radiol. 2025 Mar 30:S1076-6332(25)00206-5. doi: 10.1016/j.acra.2025.03.011. Online ahead of print.

ABSTRACT

RATIONALE AND OBJECTIVES: Coronary chronic total occlusion (CTO) and subtotal occlusion (STO) pose diagnostic challenges, differing in treatment strategies. Artificial intelligence and radiomics are promising tools for accurate discrimination. This study aimed to develop deep learning (DL) and radiomics models using coronary computed tomography angiography (CCTA) to differentiate CTO from STO lesions and compare their performance with that of the conventional method.

MATERIALS AND METHODS: CTO and STO were identified retrospectively from a tertiary hospital and served as training and validation sets for developing and validating the DL and radiomics models to distinguish CTO from STO. An external test cohort was recruited from two additional tertiary hospitals with identical eligibility criteria. All participants underwent CCTA within 1 month before invasive coronary angiography.

RESULTS: A total of 581 participants (mean age, 50 years ± 11 [SD]; 474 [81.6%] men) with 600 lesions were enrolled, including 403 CTO and 197 STO lesions. The DL and radiomics models exhibited better discrimination performance than the conventional method, with areas under the curve of 0.908 and 0.860, respectively, vs. 0.794 in the validation set (all p＜0.05), and 0.893 and 0.827, respectively, vs. 0.746 in the external test set (all p＜0.05).

CONCLUSIONS: The proposed CCTA-based DL and radiomics models achieved efficient and accurate discrimination of coronary CTO and STO.

PMID:40164533 | DOI:10.1016/j.acra.2025.03.011

Open Heart. 2025 Mar 31;12(1):e003185. doi: 10.1136/openhrt-2025-003185.

ABSTRACT

BACKGROUND: Atrial fibrillation (AF) is a common but often undiagnosed condition, increasing the risk of stroke and heart failure. Early detection is crucial, yet traditional methods struggle with AF's transient nature. This study investigates how augmenting ECG data with heart rate variability (HRV) and demographic data (age and sex) can improve AF detection.

METHODS: We analysed 35 634 12-lead ECG recordings from three public databases (China Physiological Signal Challenge-Extra, PTB-XL and Georgia), each with physician-validated AF labels. A range of convolutional neural network models, including AlexNet, VGG-16, ResNet and transformers, were tested for AF prediction, enriched with HRV and demographic data to explore the effectiveness of the multimodal approach. Each data modality (ECG, HRV and demographic) was assessed for its contribution to model performance using fivefold cross-validation. Performance improvements were evaluated across key metrics, and saliency maps were generated to provide further insights into model behaviour and identify critical features in AF detection.

RESULTS: Integrating HRV and demographic data with ECG substantially improved performance. AlexNet and VGG-16 outperformed more complex models, achieving AUROC of 0.9617 (95% CI 0.95 to 0.97) and 0.9668 (95% CI 0.96 to 0.97), respectively. Adding HRV data showed the most significant improvement in sensitivity, with AlexNet increasing from 0.9117 to 0.9225 and VGG-16 from 0.9216 to 0.9225. Combining both HRV and demographic data led to further improvements, with AlexNet achieving a sensitivity of 0.9225 (up from 0.9192 with HRV) and VGG-16 reaching 0.9113 (up from 0.9097 with HRV). The combination of HRV and demographic data resulted in the highest gains in sensitivity and area under the receiver operating characteristic curve. Saliency maps confirmed the models identified key AF features, such as the absence of the P-wave, validating the multimodal approach.

CONCLUSIONS: AlexNet and VGG-16 excelled in AF detection, with HRV data improving sensitivity, and demographic data providing additional benefits. These results highlight the potential of multimodal approaches, pending further clinical validation.

PMID:40164487 | DOI:10.1136/openhrt-2025-003185

Semin Musculoskelet Radiol. 2025 Apr;29(2):221-235. doi: 10.1055/s-0045-1802660. Epub 2025 Mar 31.

ABSTRACT

Spondyloarthritis (SpA) encompasses inflammatory disorders affecting the axial skeleton, with sacroiliitis as a hallmark feature of axial SpA (axSpA). Imaging plays a vital role in early diagnosis and disease monitoring. Magnetic resonance imaging (MRI) is the preferred modality for detecting early inflammatory changes in axSpA, whereas structural lesions are better visualized using computed tomography (CT). However, synthetic computed tomography (sCT), a technique that generates CT-like images from MRI data, including deep learning methods, zero echo time, ultrashort echo time, and gradient-recalled echo sequences, has emerged as an innovative tool. It offers detailed anatomical resolution without ionizing radiation and combines the advantages of both, MRI and CT, by enabling the simultaneous evaluation of inflammatory and structural lesions. This review explores the potential role of MRI-based sCT in assessing structural changes in the sacroiliac joints, particularly in the context of axSpA, discussing conventional imaging and highlighting the potential of sCT to enhance early detection and monitoring of sacroiliitis.

PMID:40164079 | DOI:10.1055/s-0045-1802660

Comput Biol Med. 2025 Mar 30;190:110093. doi: 10.1016/j.compbiomed.2025.110093. Online ahead of print.

ABSTRACT

BACKGROUND: Major bleeding is a severe complication in critically ill medical patients, resulting in significant morbidity, mortality, and healthcare costs. This study aims to assess the incidence and risk factors for major bleeding in hospitalised medical patients using a Natural Language Processing (NLP) model.

METHODS: We conducted a retrospective, cross-sectional observational study using electronic health records of adult patients admitted through the Emergency Department at Odense University Hospital from January 2017 to December 2022. Major bleeding during admission was identified and validated using a natural language model, with events classified according to current guidelines. Risk factors, including demographics, comorbidities, and biochemical values at admission, were evaluated. Two risk assessment models (RAMs) were developed using Cox proportional hazards regression. Validation included, bootstrapping, K-fold cross validation, and cluster analyses.

RESULTS: Of the 46,439 eligible patients, 1246 (2.7 %) experienced major bleeding. Risk factors for major bleeding included older age, male sex, alcohol consumption, higher systolic blood pressure, lower haemoglobin, and higher creatinine. RAM 1, which included biochemical data and comorbidities, demonstrated robust predictive performance (Harrell's C-statistic = 0.726). RAM 2, a simplified model without comorbidities, maintained similar predictive accuracy (C-statistic = 0.721), indicating its potential utility in clinical settings with limited resources for detailed patient histories. Results were consistent throughout validation.

CONCLUSION: This study highlights the incidence and risk factors of major bleeding in medical patients, emphasizing the predictive value of routinely measured biochemical markers. Furthermore, it shows the applicability of NLP models in identifying bleeding episodes in EHR text.

PMID:40164027 | DOI:10.1016/j.compbiomed.2025.110093