Intern Emerg Med. 2025 Jul 16. doi: 10.1007/s11739-025-04034-x. Online ahead of print.

ABSTRACT

Machine learning (ML) algorithms have the potential to enhance the prediction of adverse outcomes in patients with syncope. Recently, gradient boosting (GB) and logistic regression (LR) models have been applied to predict these outcomes following a syncope episode, using the Canadian Syncope Risk Score (CSRS) predictors. This study aims to externally validate these models and compare their performance with novel models. We included all consecutive non-low-risk patients evaluated in the emergency department for syncope between 2015 and 2017 at six Italian hospitals. The GB and LR models were trained and tested using previously validated CSRS predictors. Additionally, recently developed deep learning (TabPFN) and large language models (TabLLM) were validated on the same cohort. The area under the curve (AUC), Matthews correlation coefficient (MCC), and Brier score (BS) were compared for each model. A total of 257 patients were enrolled, with a median age of 71 years. Thirteen percent had adverse outcomes at 30 days. The GB model achieved the best performance, with an AUC of 0.78, an MCC of 0.36, and a BS of 0.42. Significant performance differences were observed compared with the TabPFN model (p < 0.01) and the TabLLM model (p = 0.01). The GB model performed only slightly better than the LR model. The predictive capability of the GB and LR models using CSRS variables was reduced when validated in an external syncope cohort characterized by a higher event rate.

PMID:40668516 | DOI:10.1007/s11739-025-04034-x

Cerebellum. 2025 Jul 16;24(5):132. doi: 10.1007/s12311-025-01887-y.

ABSTRACT

BACKGROUND AND OBJECTIVE: Degenerative cerebellar ataxia, a group of progressive neurodegenerative disorders, is characterised by cerebellar atrophy and impaired motor learning. Using CerebNet, a deep learning algorithm for cerebellar segmentation, this study investigated the relationship between cerebellar subregion volumes and motor learning ability.

METHODS: We analysed data from 37 patients with degenerative cerebellar ataxia and 18 healthy controls. Using CerebNet, we segmented four cerebellar subregions: the anterior lobe, superior posterior lobe, inferior posterior lobe, and vermis. Regression analyses examined the associations between cerebellar volumes and motor learning performance (adaptation index [AI]) and ataxia severity (Scale for Assessment and Rating of Ataxia [SARA]).

RESULTS: The inferior posterior lobe volume showed a significant positive association with AI in both single (B = 0.09; 95% CI: [0.03, 0.16]) and multiple linear regression analyses (B = 0.11; 95% CI: [0.008, 0.20]), an association that was particularly evident in the pure cerebellar ataxia subgroup. SARA scores correlated with anterior lobe, superior posterior lobe, and vermis volumes in single linear regression analyses, but these associations were not maintained in multiple linear regression analyses. This selective association suggests a specialised role for the inferior posterior lobe in motor learning processes.

CONCLUSION: This study reveals the inferior posterior lobe's distinct role in motor learning in patients with degenerative cerebellar ataxia, advancing our understanding of cerebellar function and potentially informing targeted rehabilitation approaches. Our findings highlight the value of advanced imaging technologies in understanding structure-function relationships in cerebellar disorders.

PMID:40668493 | DOI:10.1007/s12311-025-01887-y

Vis Comput Ind Biomed Art. 2025 Jul 16;8(1):18. doi: 10.1186/s42492-025-00200-2.

ABSTRACT

Unmanned aerial vehicle (UAV) tracking is a critical task in surveillance, security, and autonomous navigation applications. In this study, we propose graph neural network-tracker (GNN-tracker), a novel GNN-based UAV tracking framework that effectively integrates graph-based spatial-temporal modelling, Transformer-based feature extraction, and multi-sensor fusion to enhance tracking robustness and accuracy. Unlike traditional tracking approaches, GNN-tracker dynamically constructs a spatiotemporal graph representation, improving identity consistency and reducing tracking errors under OCC-heavy scenarios. Experimental evaluations on optical, thermal, and fused UAV datasets demonstrate the superiority of GNN-tracker (fused) over state-of-the-art methods. The proposed model achieves multiple object tracking accuracy (MOTA) scores of 91.4% (fused), 89.1% (optical), and 86.3% (thermal), surpassing TransT by 8.9% in MOTA and 7.7% in higher order tracking accuracy (HOTA). The HOTA scores of 82.3% (fused), 80.1% (optical), and 78.7% (thermal) validate its strong object association capabilities, while its frames per second of 58.9 (fused), 56.8 (optical), and 54.3 (thermal) ensures real-time performance. Additionally, ablation studies confirm the essential role of graph-based modelling and multi-sensor fusion, with performance drops of up to 8.9% in MOTA when these components are removed. Thus, GNN-tracker (fused) offers a highly accurate, robust, and efficient UAV tracking solution, effectively addressing real-world challenges across diverse environmental conditions and multiple sensor modalities.

PMID:40668492 | DOI:10.1186/s42492-025-00200-2

J Comput Aided Mol Des. 2025 Jul 16;39(1):53. doi: 10.1007/s10822-025-00637-w.

ABSTRACT

Alzheimer's disease (AD) is a progressive neurodegenerative disorder lacking effective therapies. Glycogen synthase kinase-3β (GSK-3β), a key regulator of Aβ aggregation and Tau hyperphosphorylation, has emerged as a promising therapeutic target. Here, we present a novel two-stage virtual screening (VS) framework that integrates an interpretable random forest (RF) model (AUC = 0.99) with a deep learning-based molecular docking platform, KarmaDock (NEF0.5% = 1.0), to identify potential GSK-3β inhibitors from natural products. The model's interpretability was enhanced using SHAP analysis to uncover key fingerprint features driving activity predictions. A curated natural compound library (n = 25,000) from TCMBank and HERB was constructed under drug-likeness constraints, and validated using multi-level decoy sets. Three compounds derived from Clausena and Psoralea exhibited favorable pharmacokinetic profiles in silico, including blood-brain barrier permeability and low neurotoxicity. Molecular docking, pharmacophore modeling, and molecular dynamics simulations confirmed their stable interactions with critical GSK-3β binding sites. Notably, our approach combines explainability and deep learning to enhance screening accuracy and interpretability, addressing limitations in traditional black-box models. While current findings are computational, they offer theoretical support and provide actionable leads for future experimental validation of natural GSK-3β inhibitors.

PMID:40668407 | DOI:10.1007/s10822-025-00637-w

Neuroradiology. 2025 Jul 16. doi: 10.1007/s00234-025-03691-1. Online ahead of print.

ABSTRACT

PURPOSE: The relationship between mild neurovascular conflict (NVC) and trigeminal neuralgia (TN) remains ill-defined, especially as mild NVC is often seen in asymptomatic population without any facial pain. We aim to analyze the trigeminal nerve microstructure using artificial intelligence (AI) to distinguish symptomatic and asymptomatic nerves between idiopathic TN (iTN) and the asymptomatic control group with incidental grade‑1 NVC.

METHODS: Seventy-eight symptomatic trigeminal nerves with grade-1 NVC in iTN patients, and an asymptomatic control group consisting of Bell's palsy patients free from facial pain (91 grade-1 NVC and 91 grade-0 NVC), were included in the study. Three hundred seventy-eight radiomic features were extracted from the original MRI images and processed with Laplacian-of-Gaussian filters. The dataset was split into 80% training/validation and 20% testing. Nested cross-validation was employed on the training/validation set for feature selection and model optimization. Furthermore, using the same pipeline approach, two customized deep learning models, Dense Atrous Spatial Pyramid Pooling (ASPP) -201 and MobileASPPV2, were classified using the same pipeline approach, incorporating ASPP blocks.

RESULTS: Performance was assessed over ten and five runs for radiomics-based and deep learning-based models. Subspace Discriminant Ensemble Learning (SDEL) attained an accuracy of 78.8%±7.13%, Support Vector Machines (SVM) reached 74.8%±9.2%, and K-nearest neighbors (KNN) achieved 79%±6.55%. Meanwhile, DenseASPP-201 recorded an accuracy of 82.0 ± 8.4%, and MobileASPPV2 achieved 73.2 ± 5.59%.

CONCLUSION: The AI effectively distinguished symptomatic and asymptomatic nerves with grade‑1 NVC. Further studies are required to fully elucidate the impact of vascular and nonvascular etiologies that may lead to iTN.

PMID:40668403 | DOI:10.1007/s00234-025-03691-1

Int J Oral Maxillofac Implants. 2025 Jul 16;0(0):1-20. doi: 10.11607/jomi.11415. Online ahead of print.

ABSTRACT

PURPOSE: To develop auxiliary devices for intraoral (IO) scanning of complete-arch implants using a deep-learning AI model.

MATERIALS AND METHODS: A total of 338 sets of 3D imaging data were collected from a dental laboratory. Of these, 300 sets of complete dental arches were used for training, 38 sets for validation, and 10 edentulous arches with 4-6 dental implants for testing. Auxiliary devices, with landmarks placed between implants to aid in image stitching, were manually designed and used as a control. A Multi-Layer Perceptron artificial neural network was employed to predict the positions of the landmarks, using normalized implant coordinates as input and landmark coordinates as output. The model was validated and evaluated using the test set to assess the fit of the base and the surface area of the landmarks.

RESULTS: The bounding box loss for the training and validation sets converged to 0.02 and 0.01, respectively, indicating high precision in predicting landmark positions. The objectness loss stabilized at 0.05 for the training set and 0.03 for the validation set, confirming the model's robust detection capability. The root mean square (RMS) of the device base was 0.117 ± 0.053 mm, significantly smaller than the clinical threshold of 0.300 mm (p < 0.001). The surface area of the AI-generated device landmarks (762.0 ± 141.7 mm²) was significantly smaller than that of the manually designed control (1307.1 ± 286.1 mm², p = 0.001).

CONCLUSIONS: The AI model demonstrates exceptional performance in the task. The base of the AI-generated auxiliary device fits well with the edentulous region, while its landmark teeth are smaller than those of the manually designed control.

PMID:40668357 | DOI:10.11607/jomi.11415

bioRxiv [Preprint]. 2025 Jun 27:2025.06.22.660948. doi: 10.1101/2025.06.22.660948.

ABSTRACT

All genes are not equal before literature. Despite the explosion of genomic data, a significant proportion of human protein-coding genes remain poorly characterized ("ghost genes"). Due to sociological dynamics in research, scientific literature disproportionately focuses on already well-annotated genes, reinforcing existing biases (bandwagon effect). This literature bias often permeates machine learning (ML) models trained on gene annotation tasks, leading to predictions that favor well-studied genes. Consequently, standard ML performance metrics may overestimate biological relevance by overfitting literature-derived patterns. To address this challenge, we developed GhostBuster, an encoder-decoder ML platform designed to predict gene functions, disease associations and interactions while minimizing literature bias. We first compared the impact of biased (Gene Ontology) versus unbiased training datasets (LINCS, TCGA, STRING). While literature-biased sources yielded higher ML metrics, they also amplified bias by prioritizing well-characterized genes. In contrast, models trained on unbiased datasets were 2-3× more effective at identifying recently discovered gene annotations. Notably, one of the unbiased channels (TCGA), combined minimal amounts of literature bias with robust performance, at a test ROC-AUC of 0.8-0.95. We demonstrate that GhostBuster can be applied to predict novel gene functions, refine pathway memberships, and prioritize intergenic GWAS hits. As the first ML framework explicitly designed to counteract literature bias, GhostBuster offers a powerful tool for uncovering the roles of understudied genes in cellular function, disease, and molecular networks.

PMID:40667329 | PMC:PMC12262676 | DOI:10.1101/2025.06.22.660948

bioRxiv [Preprint]. 2025 Jun 27:2025.06.23.661188. doi: 10.1101/2025.06.23.661188.

ABSTRACT

Oncogene amplification is a key driver of cancer pathogenesis and is often mediated by extrachromosomal DNA (ecDNA). EcDNA amplifications are associated with increased pathogenicity of cancer and poorer outcomes for patients. EcDNA can be detected accurately using fluorescence in situ hybridization (FISH) when cells are arrested in metaphase. However, the majority of cancer cells are non-mitotic and must be analyzed in interphase, where it is difficult to discern extrachromosomal amplifications from chromosomal amplifications. Thus, there is a need for methods that accurately predict oncogene amplification status from interphase cells. Here, we present interSeg, a deep learning-based tool to cytogenetically determine the amplification status as EC-amp, HSR-amp, or not amplified from interphase FISH images. We trained and validated interSeg on 652 images (40,446 nuclei). Tests on 215 cultured cell and tissue model images (9,733 nuclei) showed 89% and 97% accuracy at the nuclear and sample levels, respectively. The neuroblastoma patient tissue hold-out set (67 samples and 1,937 nuclei) also revealed 97% accuracy at the sample level in detecting the presence of focal amplification. In experimentally and computationally mixed images, interSeg accurately predicted the level of heterogeneity. The results showcase interSeg as an important method for analyzing oncogene amplifications.

PMID:40667255 | PMC:PMC12262287 | DOI:10.1101/2025.06.23.661188

bioRxiv [Preprint]. 2025 Jun 17:2025.06.11.659181. doi: 10.1101/2025.06.11.659181.

ABSTRACT

Regulator of G protein Signaling-14 (RGS14), an intracellular inactivator of G protein-coupled receptor (GPCR) signaling, has long been considered an undruggable protein due to its shallow and relatively featureless protein- protein interaction interface. Here, we describe the successful identification and validation of a tractable chemotype that selectively inhibits the GTPase-accelerating protein (GAP) activity of RGS14. Combining structure-guided virtual screening, ligand docking across multiple available receptor conformers, and enrichment validation, we progressed from an initial first-generation active compound, Z90276197, to over 40 second-generation active analogs with improved potency. These inhibitors are predicted to engage a conserved, solvent-exposed "canyon" in the RGS14 RGS-box, which interacts with the Gα switch I region. Predicted binding poses underscored the importance of non-polar interactions and shape complementarity over polar interactions in engaging RGS14's shallow Gα-binding canyon and revealed a recurring "ambidextrous" pattern of substituent orientations. Functional GAP inhibition was confirmed in fluorescence-based and the gold-standard radioactive GTP hydrolysis assays. Two second-generation analogs, Z55660043 and Z55627844, inhibited RGS14 GAP activity in both assays and without measurable cytotoxicity. Deep learning-based scoring of predicted docking poses further supported observed affinity gains from methyl-ester additions. One analog demonstrated favorable in vivo pharmacokinetics and CNS penetration. Collectively, our findings establish an example of tractable small molecule inhibition of a regulatory interface of a G protein and illustrate how machine learning-enhanced docking can guide ligand optimization for shallow protein surfaces. This work opens the door to future development of RGS14 inhibitors as potential therapeutics for central nervous system and metabolic disorders.

PMID:40667230 | PMC:PMC12262430 | DOI:10.1101/2025.06.11.659181

bioRxiv [Preprint]. 2025 Jun 25:2025.06.19.660198. doi: 10.1101/2025.06.19.660198.

ABSTRACT

Advances in high-throughput imaging and experimental automation have dramatically increased the scale of biological datasets, creating a growing need for tools that can efficiently identify and localize features in complex image data. Although deep learning has transformed image analysis, methods such as region-based convolutional neural networks remain underutilized in biology due to technical barriers such as coding requirements and reliance on cloud infrastructure. We present EZ-FRCNN, a locally hosted, user-friendly package that enables the accessible and scalable application of object detection to biological datasets. Through graphical and scriptable interfaces, users can annotate data, train models, and perform inference entirely offline. We demonstrate its utility in detecting cell phenotypes for large-scale screening, enabling the first label-free tracking of grinder motion in freely moving C. elegans to quantify feeding dynamics, and identifying animals in naturalistic environments for ecological field studies. These once-infeasible analyses now enable rapid screening of cell therapies, investigation of internal state-behavior coupling without immobilization or genetic modification, and efficient wildlife tracking with minimal computational cost. Together, these examples demonstrate how accessible tools like EZ-FRCNN can drive new biological discoveries in both laboratory and field environments.

PMID:40667197 | PMC:PMC12262326 | DOI:10.1101/2025.06.19.660198

bioRxiv [Preprint]. 2025 Jun 27:2025.06.26.661760. doi: 10.1101/2025.06.26.661760.

ABSTRACT

Malaria in sub-Saharan Africa is transmitted by mosquitoes, in particular the Anopheles gambiae complex. Efforts to control the spread of malaria have often focused on these vectors, but relatively little is known about the relationships between populations and species in the Anopheles complex. Here, we first quantify the genetic structure of mosquito populations in sub-Saharan Africa using unsupervised machine learning. We then adapt and apply an innovative generative deep learning algorithm to infer the joint evolutionary history of populations sampled in Guinea and Burkina Faso, West Africa. We further develop a novel model selection approach and discover that an evolutionary model with migration fits this pair of populations better than a model without post-split migration. For the migration model, we find that our method outperforms earlier work based on summary statistics, especially in capturing population genetic differentiation. These findings demonstrate that machine learning and generative models are a valuable direction for future understanding of the evolution of malaria vectors, including the joint inference of demography and natural selection. Understanding changes in population size, migration patterns, and adaptation in hosts, vectors, and pathogens will assist malaria control interventions, with the ultimate goal of predicting nuanced outcomes from insecticide resistance to population collapse.

PMID:40667127 | PMC:PMC12262376 | DOI:10.1101/2025.06.26.661760

bioRxiv [Preprint]. 2025 Jun 18:2025.03.21.644663. doi: 10.1101/2025.03.21.644663.

ABSTRACT

Cognitive modeling in psychology and neuroscience provides a formal approach to formulate and test hypotheses of cognitive processes. Such models rely on cognitive parameters that are intended to represent interpretable and identifiable psychological constructs. However, accurately and reliably estimating model parameters remains challenging due to common issues such as limited data, measurement noise, experimental constraints, and model complexity, hindering the interpretability and identifiability of cognitive parameters. Given the recent success of advanced optimization methods in deep learning, we investigate whether a deep learning pipeline that integrates neural networks and modern optimization techniques can improve parameter estimation in reinforcement learning (RL) models. We compare this approach with the Nelder-Mead method ( fminsearch ), the de facto optimization approach in cognitive modeling, by fitting RL models to ten diverse value-based decision-making datasets collected from both humans and animals. Surprisingly, while both approaches achieve comparable predictive performance, they produce distinct parameter estimates, indicating fitting performance alone is insufficient in identifying these parameters. We thus systematically evaluate the reliability of parameters estimated by each approach and find that parameters estimated via the deep learning pipeline consistently demonstrate smaller gaps between training and test performance (better generalizability), increased resistance to parameter perturbations (enhanced robustness), improved recovery of ground-truth parameters in low-data regimes (stronger identifiability), and greater consistency across repeated measurements from the same individuals (better test-retest reliability). Our findings advocate for the deep learning pipeline and systematic evaluation of cognitive parameters to better link these parameters to psychological constructs and neural mechanisms.

PMID:40666915 | PMC:PMC12262606 | DOI:10.1101/2025.03.21.644663

bioRxiv [Preprint]. 2025 Jun 17:2025.06.17.659969. doi: 10.1101/2025.06.17.659969.

ABSTRACT

CRISPR-associated transposons (CAST) enable programmable, RNA-guided DNA integration, marking a transformative advancement in genome engineering. A central player in the type V-K CAST system is the AAA+ ATPase TnsC, which assembles into helical filaments on double-stranded DNA (dsDNA) to orchestrate target site recognition and transposition. Despite its essential role, the molecular mechanisms underlying TnsC filament nucleation and elongation remain poorly understood. Here, multiple-microsecond and free energy simulations are combined with deep learning-based Graph Attention Network (GAT) models to elucidate the mechanistic principles of TnsC filament formation and growth. Our findings reveal that ATP binding promotes TnsC nucleation by inducing DNA remodelling and stabilizing key protein-DNA interactions, particularly through conserved residues in the initiator-specific motif (ISM). Furthermore, GNN-based attention analyses identify a directional bias in filament elongation in the 5'→3' direction and uncover a dynamic compensation mechanism between incoming and bound monomers that facilitate directional growth along dsDNA. By leveraging deep learning-based graph representations, our GAT model provides interpretable mechanistic insights from complex molecular simulations and is readily adaptable to a wide range of biological systems. Altogether, these findings establish a mechanistic framework for TnsC filament dynamics and directional elongation, advancing the rational design of CAST systems with enhanced precision and efficiency.

PMID:40666904 | PMC:PMC12262710 | DOI:10.1101/2025.06.17.659969

bioRxiv [Preprint]. 2025 Jun 27:2025.06.24.661337. doi: 10.1101/2025.06.24.661337.

ABSTRACT

Probabilistic models such as the sequentially Markovian coalescent (SMC) have long provided a powerful framework for population genetic inference, enabling reconstruction of demographic history and ancestral relationships from genomic data. However, these methods are inherently specialized, relying on predefined assumptions and/or limited scalability. Recent advances in simulation and deep learning provide an alternative approach: learning directly to generalize from synthetic genetic data to infer specific hidden evolutionary processes. Here we reframe the inference of coalescence times as a problem of translation between two biological languages: the sparse, observable patterns of mutation along the genome and the unobservable ancestral recombination graph (ARG) that gave rise to them. Inspired by large language models, we develop cxt, a decoder-only transformer that autoregressively predicts coalescent events conditioned on local mutational context. We show that cxt performs on par with state-of-the-art MCMC-based likelihood models across a broad range of demographic scenarios, including both in-distribution and out-of-distribution settings. Trained on simulations spanning the stdpopsim catalog, the model generalizes robustly and enables efficient inference at scale, producing over a million coalescence predictions in minutes. In addition cxt produces a well calibrated approximate posterior distribution of its predictions, enabling principled uncertainty quantification. Our work moves towards a foundation model for population genetics, bridging deep learning and coalescent theory to enable flexible, scalable inference of genealogical history from genomic data.

PMID:40666889 | PMC:PMC12262695 | DOI:10.1101/2025.06.24.661337

bioRxiv [Preprint]. 2025 Jun 27:2025.05.26.656232. doi: 10.1101/2025.05.26.656232.

ABSTRACT

Microbial communities are essential for sustaining ecosystem functions in diverse environments, including the human gut. Phages interact dynamically with their prokaryotic hosts and play a crucial role in shaping the structure and function of microbial communities. Previous approaches for inferring phage-host interactions (PHIs) from metagenomic data are constrained by low sensitivity and the inability to accurately capture ecological relationships. To overcome these limitations, we developed PHILM ( P hage- H ost Interaction L earning from M etagenomic profiles), a deep learning framework that predicts PHIs directly from the taxonomic profiles of metagenomic data. We validated PHILM on both synthetic datasets generated by ecological models and real-world data, finding that it consistently outperformed the co-abundance-based approach for inferring PHIs. When applied to a large-scale metagenomic dataset comprising 7,016 stool samples from healthy individuals, PHILM identified 90% more genus-level PHIs than the traditional assembly-based approach. In a longitudinal dataset tracking PHI dynamics, PHILM's latent representations recapitulated microbial succession patterns originally described using taxonomic abundances. Furthermore, we demonstrated that PHILM's latent representations served as more discriminative features than taxonomic abundance-based features for disease classifications. In summary, PHILM represents a novel computational framework for predicting phage-host interactions from metagenomic data, offering valuable insights for both microbiome science and translational medicine.

PMID:40666868 | PMC:PMC12262735 | DOI:10.1101/2025.05.26.656232

bioRxiv [Preprint]. 2025 Jun 27:2025.06.22.660946. doi: 10.1101/2025.06.22.660946.

ABSTRACT

The correlation between transcriptomic (Tx) and proteomic (Px) profiles remains modest, typically around r = 0.5 across genes and r = 0.3 across samples, limiting the utility of transcriptomic data as a proxy for protein abundance. To address this, we introduce Proteomizer, a deep learning platform designed to infer a sample's Px landscape from its Tx and miRNomic (Mx) profiles. Trained on 8,613 matched Tx-Mx-Px samples from TCGA and CPTAC, Proteomizer achieved a Tx-Px correlation of r = 0.68, representing the highest performance reported to date for this task. We further developed a Monte Carlo simulation framework to evaluate the impact of proteomization on differential expression analysis. Proteomizer substantially improved the accuracy of differential gene expression detection, with p-value precision increasing by up to 62-fold, and by as much as six orders of magnitude for a subset of genes enriched in mitochondrial and ribosomal functions. However, performance gains did not generalize to unseen tissue types or datasets generated using different protocols. Finally, we applied explainable AI (XAI) techniques to identify regulatory relations contributing to Tx-Px discrepancies. Our predictions from 100 highly annotated genes were cross-compared against by a literature-based biological knowledge graph of 322 million annotations: our explainers achieved a ROC-AUC of 0.74 in predicting miRNA-gene downregulation interactions. To our knowledge, this is the first study to systematically evaluate the biological relevance, limitations, and interpretability of proteomization models, establishing Proteomizer as a state-of-the-art tool for multiomic integration and hypothesis generation.

PMID:40666834 | PMC:PMC12262288 | DOI:10.1101/2025.06.22.660946

Med Phys. 2025 Jul;52(7):e18015. doi: 10.1002/mp.18015.

ABSTRACT

BACKGROUND: Real-time image-guided radiation therapy (IGRT) was first clinically implemented more than 25 years ago but is yet to find widespread adoption. Existing approaches to real-time IGRT require dedicated or specialized equipment that is not available in most treatment centers and most techniques focus exclusively on targets without tracking the surrounding organs-at-risk (OARs).

PURPOSE: To address the need for inexpensive real-time IGRT, we developed Voxelmap, a deep learning framework that achieves 3D respiratory motion estimation and volumetric imaging using the data and resources already available in standard clinical settings. This framework can also be adapted to other imaging modalities such as MRI-Linacs. In contrast with existing approaches, which constrain the solution space with linear priors, Voxelmap encourages diffeomorphic mappings that are topology-preserving and invertible.

METHODS: Deformable image registration and forward-projection or slice extraction were used to generate patient-specific training datasets of 3D deformation vector fields (DVFs) and 2D images (or k-space data) from pretreatment 4D-CT or 4D-MRI scans. The XCAT and CoMBAT digital phantoms and SPARE Grand Challenge Dataset provided synthetic and patient data, respectively. Five network architectures were used to predict 3D DVFs from 2D imaging data. Networks A-C were trained on x-ray images, Network D was trained on MR images and Network E was trained on k-space data. Using Voxelmap, network-generated 3D DVFs were used to warp both structures contoured on the peak-exhale pretreatment image and the image itself to enable simultaneous target and OAR tracking and volumetric imaging. Using the standard-of-care approach, contours were expanded to internal target volumes.

RESULTS: Validating on digital phantom data for x-ray guided treatments of cardiac arrhythmia, mean Dice similarity between predicted and ground-truth target and OAR contours for Networks A-C ranged from 0.81 ± 0.05 to 0.82 ± 0.05 and 0.78 ± 0.04 to 0.81 ± 0.04, respectively, while target centroid error ranged from 2.0 ± 0.5 to 2.3 ± 0.9 mm. For MRI-based digital phantom data, mean Dice similarity for target and OAR contours was 0.91 ± 0.06 and 0.90 ± 0.02 for both Networks D and E, while target centroid error ranged from 1.7 ± 0.8 to 1.8 ± 0.8 mm. For x-ray-based lung cancer patient data, mean Dice similarity for target and OAR contours for Networks A-C ranged from 0.86 ± 0.05 to 0.89 ± 0.04 and 0.94 ± 0.01 to 0.97 ± 0.01, respectively. However, in terms of target centroid error, only Network A outperformed an ITV-based approach at 1.8 ± 0.7 mm while Networks B and C exhibited large errors of 2.7 ± 1.2 to 3.5 ± 1.4 mm, respectively. Target volumes dynamically shifted using Voxelmap were 31 % smaller than the standard-of-care.

CONCLUSIONS: Voxelmap provides a generalized, open-source tool for intrafraction respiratory motion monitoring and volumetric imaging. Comparing tracking errors across synthetic and patient data revealed that certain network architectures are more robust to the scatter and noise profiles encountered in typical clinical settings. These learnings will inform future developments in real-time motion tracking. Our code is available at https://github.com/Image-X-Institute/Voxelmap .

PMID:40665474 | DOI:10.1002/mp.18015

BMC Med Inform Decis Mak. 2025 Jul 15;25(1):264. doi: 10.1186/s12911-025-03110-8.

ABSTRACT

PURPOSE: Accurate identification of bone marrow invasion (BMI) is critical for determining the prognosis of and treatment strategies for lymphoma. Although bone marrow biopsy (BMB) is the current gold standard, its invasive nature and sampling errors highlight the necessity for noninvasive alternatives. We aimed to develop and validate an interpretable machine learning model that integrates clinical data, 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) parameters, radiomic features, and deep learning features to predict BMI in lymphoma patients.

METHODS: We included 159 newly diagnosed lymphoma patients (118 from Center I and 41 from Center II), excluding those with prior treatments, incomplete data, or under 18 years of age. Data from Center I were randomly allocated to training (n = 94) and internal test (n = 24) sets; Center II served as an external validation set (n = 41). Clinical parameters, PET/CT features, radiomic characteristics, and deep learning features were comprehensively analyzed and integrated into machine learning models. Model interpretability was elucidated via Shapley Additive exPlanations (SHAPs). Additionally, a comparative diagnostic study evaluated reader performance with and without model assistance.

RESULTS: BMI was confirmed in 70 (44%) patients. The key clinical predictors included B symptoms and platelet count. Among the tested models, the ExtraTrees classifier achieved the best performance. For external validation, the combined model (clinical + PET/CT + radiomics + deep learning) achieved an area under the receiver operating characteristic curve (AUC) of 0.886, outperforming models that use only clinical (AUC 0.798), radiomic (AUC 0.708), or deep learning features (AUC 0.662). SHAP analysis revealed that PET radiomic features (especially PET_lbp_3D_m1_glcm_DependenceEntropy), platelet count, and B symptoms were significant predictors of BMI. Model assistance significantly enhanced junior reader performance (AUC improved from 0.663 to 0.818, p = 0.03) and improved senior reader accuracy, although not significantly (AUC 0.768 to 0.867, p = 0.10).

CONCLUSION: Our interpretable machine learning model, which integrates clinical, imaging, radiomic, and deep learning features, demonstrated robust BMI prediction performance and notably enhanced physician diagnostic accuracy. These findings underscore the clinical potential of interpretable AI to complement medical expertise and potentially reduce the reliance on invasive BMB for lymphoma staging.

PMID:40665334 | DOI:10.1186/s12911-025-03110-8

BMC Med Inform Decis Mak. 2025 Jul 15;25(1):268. doi: 10.1186/s12911-025-03062-z.

ABSTRACT

BACKGROUND: Accurate delineation of organs at risk (OARs) is crucial for precision radiotherapy. Most previous autosegmentation models were only constructed for single anatomical region without evaluation of dosimetric impact. We aimed to validate the clinical practicability of deep-learning (DL) models for autosegmentation of whole-body OARs with respect to delineation accuracy, clinical acceptance and dosimetric impact.

METHODS: OARs in various anatomical regions, including the head and neck, thorax, abdomen, and pelvis, were automatedly delineated by DL models (DLD) and compared to manual delineations (MD) by an experienced radiation oncologist (RO). The geometric performance was evaluated using the Dice similarity coefficient (DSC) and average surface distance (ASD). RO A corrected DLD to create delineations approved in clinical practice (CPD). RO B graded the accuracy of DLD to assess clinical acceptance. The dosimetric impact was determined by assessing the difference in dosimetric parameters for each OAR in the DLD-based radiotherapy plan (Plan_DLD) and the CPD-based radiotherapy plan (Plan_CPD).

RESULTS: The automatic delineation model has a high OAR delineation accuracy, and the median DSCs can reach 0.841 (IQR, 0.791-0.867) in the head and neck OAR, 0.903 (IQR, 0.777-0.932) in thoracic OAR, 0.847 (IQR, 0.834-0.931) in abdominal OAR, 0.916 (IQR, 0.906-0.964) in pelvic OAR. The majority of DL-generated OARs were graded as clinically acceptable with no editing or little editing needed. No significant differences in dosimetric parameters were found by comparing Plan_DLD with Plan_CPD.

CONCLUSIONS: For OARs of whole bodily regions, DL-based segmentation is fast; DL models perform sufficiently well for clinical practice with respect to delineation accuracy, clinical accepatance and dosimetric impact.

PMID:40665308 | DOI:10.1186/s12911-025-03062-z

BMC Musculoskelet Disord. 2025 Jul 15;26(1):682. doi: 10.1186/s12891-025-08842-2.

ABSTRACT

BACKGROUND: Currently, the application of convolutional neural networks (CNNs) in artificial intelligence (AI) for medical imaging diagnosis has emerged as a highly promising tool. In particular, AI-assisted diagnosis holds significant potential for orthopedic and emergency department physicians by improving diagnostic efficiency and enhancing the overall patient experience. This systematic review and meta-analysis has the objective of assessing the application of AI in diagnosing facial fractures and evaluating its diagnostic performance.

METHODS: This study adhered to the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) and PRISMA-Diagnostic Test Accuracy (PRISMA-DTA). A comprehensive literature search was conducted in the PubMed, Cochrane Library, and Web of Science databases to identify original articles published up to December 2024. The risk of bias and applicability of the included studies were assessed using the QUADAS-2 tool. The results were analyzed using a Summary Receiver Operating Characteristic (SROC) curve.

RESULTS: A total of 16 studies were included in the analysis, with contingency tables extracted from 11 of them. The pooled sensitivity was 0.889 (95% CI: 0.844-0.922), and the pooled specificity was 0.888 (95% CI: 0.834-0.926). The area under the Summary Receiver Operating Characteristic (SROC) curve was 0.911. In the subgroup analysis of nasal and mandibular fractures, the pooled sensitivity for nasal fractures was 0.851 (95% CI: 0.806-0.887), and the pooled specificity was 0.883 (95% CI: 0.862-0.902). For mandibular fractures, the pooled sensitivity was 0.905 (95% CI: 0.836-0.947), and the pooled specificity was 0.895 (95% CI: 0.824-0.940).

CONCLUSIONS: AI can be developed as an auxiliary tool to assist clinicians in diagnosing facial fractures. The results demonstrate high overall sensitivity and specificity, along with a robust performance reflected by the high area under the SROC curve.

CLINICAL TRIAL NUMBER: This study has been prospectively registered on Prospero, ID:CRD42024618650, Creat Date:10 Dec 2024. https://www.crd.york.ac.uk/PROSPERO/view/CRD42024618650 .

PMID:40665293 | DOI:10.1186/s12891-025-08842-2