Computer Methods And Programs In Biomedicine[JOURNAL]
Sun
200 papers
Predicting intrinsic clearance using deep learning-based drug-metabolic enzyme interaction features on an IVIVE-harmonized endpoint dataset.
Lee H, Kang H, Chae JW … +4 more , Yun HY, Goo S, Namgoong H, Jung S
Comput Methods Programs Biomed · 2026 Sep · PMID 42177842 · Publisher ↗
BACKGROUND AND OBJECTIVE: Intrinsic clearance is a key pharmacokinetic parameter in drug development because it influences systemic exposure and dose selection. Conventional in vitro-in vivo extrapolation (IVIVE) approac...
BACKGROUND AND OBJECTIVE: Intrinsic clearance is a key pharmacokinetic parameter in drug development because it influences systemic exposure and dose selection. Conventional in vitro-in vivo extrapolation (IVIVE) approaches are useful but often require additional experimental resources and may be difficult to apply in the earliest stages of compound screening. This study aimed to develop a biologically informed computational framework for intrinsic clearance prediction by integrating drug-target interaction (DTI)-derived features from 23 hepatic metabolic proteins with physicochemical properties, using an IVIVE-based endpoint harmonization approach to unify heterogeneous endpoint labels. METHODS: A multi-source human clearance dataset was assembled and harmonized into a common intrinsic clearance endpoint. A pretrained DTI model based on ChemBERTa and ProtBERT was used to generate precomputed interaction features for CYP, UGT, and SULT family proteins. These features were combined with compound structure and physicochemical descriptors in downstream clearance prediction models using either a multilayer perceptron (MLP) or a transformer encoder. Internal evaluation included repeated random 5-fold cross-validation and scaffold-split analysis, and an independent external evaluation set of 185 compounds was additionally examined. RESULTS: Feature settings incorporating DTI-derived information showed improved regression performance compared with the descriptor-only baseline across most evaluated metrics. Under the augmented setting, the MLP model with DTI + LogP/Fup showed the highest overall regression performance among the evaluated models (r = 0.2505 ± 0.0792, r = 0.2364 ± 0.0879), while the highest CI was observed in the transformer encoder model with DTI + LogP/Fup (0.6915 ± 0.0144). On the external evaluation set, the MLP model showed an AFE of 1.116, with 52.970% and 70.792% of predictions falling within 2-fold and 3-fold error, respectively. Comparisons with published IVIVE-related studies should be interpreted as contextual rather than direct head-to-head validation. CONCLUSIONS: The proposed framework supports intrinsic clearance estimation by combining biologically informed DTI-derived features with physicochemical information in an interpretable prediction setting. Although the predictive performance remains modest, the results suggest potential utility as a supportive early-stage prioritization tool when standardized experimental clearance data are limited. Further work is needed to expand the dataset, strengthen external validation with independent cohorts, and incorporate additional elimination pathways such as renal clearance and transporter-mediated disposition.
Liver point cloud registration via multi-feature fusion and hyperbolic embedding for augmented reality surgical navigation.
Yang X, He B, Dai Y … +3 more , Luo H, Wang L, Jia F
Comput Methods Programs Biomed · 2026 Sep · PMID 42177841 · Publisher ↗
BACKGROUND AND OBJECTIVE: Augmented reality surgical navigation can help surgeons precisely locate tumors, veins, arteries and other critical structures during laparoscopic liver surgery. A key step in the navigation is...
BACKGROUND AND OBJECTIVE: Augmented reality surgical navigation can help surgeons precisely locate tumors, veins, arteries and other critical structures during laparoscopic liver surgery. A key step in the navigation is registering the preoperative 3D liver model with the intraoperative liver surface. Accurate registration first requires rigid alignment; however, rigid point cloud registration in laparoscopic scenes remains a challenging task. METHODS: We propose a method that employs a dual-backbone to extract features from the point cloud, fusing these features to obtain robust descriptors. The fused features are then mapped to hyperbolic space, where hyperbolic attention is introduced to enhance feature representation further. In addition, position encoding is incorporated to improve the model's geometric perception. Following this, a coarse-to-fine matching strategy is applied to perform patch-level and point-level matching sequentially. Finally, the transformation matrix is computed using RANSAC based on the derived point correspondences. RESULTS: We evaluated our method on real-world and simulated datasets, including porcine data and two valuable clinical patient datasets acquired under pneumoperitoneum. The results show strong performance in the rigid registration stage. In several settings, the method achieves state-of-the-art results. On the 3Dircadb01 simulated dataset, our method achieved the highest inlier ratio of 64.53% and the lowest surface registration error of 4.07mm. CONCLUSIONS: Our proposed method improves rigid registration accuracy and robustness for the initial alignment from preoperative medical image space to intraoperative laparoscopic image space in AR surgical navigation. It provides a reliable rigid initialization for subsequent non-rigid registration. Our code will be released at Github.
CardioRadNet: Cardiac mass diagnosis through integrated segmentation and radiomic analysis.
Ferretti M, Pagliaccia M, Baggiano A … +7 more , Lovato L, Angeli F, Armillotta M, Bergamaschi L, Pizzi C, Pontone G, Corino VDA
Comput Methods Programs Biomed · 2026 Sep · PMID 42176411 · Publisher ↗
BACKGROUND AND OBJECTIVE: Cardiac masses (CMs), though rare, include a wide spectrum of benign and malignant lesions that require distinct therapeutic strategies. Prior studies typically address narrow classification tas...
BACKGROUND AND OBJECTIVE: Cardiac masses (CMs), though rare, include a wide spectrum of benign and malignant lesions that require distinct therapeutic strategies. Prior studies typically address narrow classification tasks (e.g., thrombus vs tumor) and rely on manually delineated regions. Building on these gaps, we introduce CardioRadNet: the first integrated framework combining deep learning segmentation and radiomics-based classification on contrast-free T1-weighted cardiac MR, designed to differentiate benign from malignant CMs. METHODS: A total of 127 patients with pathologically confirmed CMs (62 malignant, 65 benign) were included. A segmentation network incorporating point-based guidance was developed for mass delineation. Radiomic features were extracted from both manually and semiautomatically segmented volumes, and two separate radiomics-based classification models were developed accordingly. Feature selection and classifier performance were optimized using a 5-fold cross-validation. RESULTS: The segmentation network achieved a Dice score of 0.78, with 88% of the radiomic features extracted from the semiautomatic ROIs showing good reproducibility (ICC > 0.6) when compared with those derived from manual ROIs. The two models achieved identical balanced accuracy (0.85) and the same number of misclassifications (both used 10 features). Notably, the semiautomated ROI model yielded one fewer false negative, thereby reducing missed malignant cases. CONCLUSION: CardioRadNet offers a novel, accurate and contrast-free solution for comprehensive CM classification. Unlike prior studies, it includes the full spectrum of CMs and uses semiautomated segmentation for broader clinical applicability. Overall, this approach supports scalable integration into routine CMR workflows as a decision-support tool for early risk stratification and, in turn, improved patient management.
Physics-informed DynUNet for brain metastasis segmentation.
Güzel M, Baykan ÖK
Comput Methods Programs Biomed · 2026 Aug · PMID 42172695 · Publisher ↗
BACKGROUND: In neuro-oncology, detecting, segmenting, and delineating the boundaries of small-volume brain metastatic foci remains a significant challenge. The lack of explicit biological information on metastasis growth...
BACKGROUND: In neuro-oncology, detecting, segmenting, and delineating the boundaries of small-volume brain metastatic foci remains a significant challenge. The lack of explicit biological information on metastasis growth and spread in standard deep learning architectures further limits low-volume metastatic lesions. This study investigates whether integrating physics-informed (PI) tumor growth models into segmentation architectures can overcome these size-dependent limitations. METHODS: Using the BraTS-METS 2023 dataset, we integrated a physics-based growth model with DynUNet to construct PI-DynUNet and compared it with three U-Net variants under controlled conditions. All models were trained on the same data without data augmentation, using matched parameter counts, identical hyperparameters, and deterministic settings. We compared seven physics regularization weights (λ) with 5-fold cross-validation and evaluated performance in six lesion-size categories using Dice, IoU and HD95. To assess clinical context-specific performance, we calculated scenario-weighted Dice coefficients for RANO progression assessment, radiotherapy planning, and surgical decision-making. RESULTS: PI-DynUNet achieved effective metastasis segmentation across all BraTS regions. Relative to baseline DynUNet, it improved whole tumor (WT) Dice by 1.8 %, tumor core (TC) Dice by 2.5 %, and enhancing tumor (ET) Dice by 2.6 %. For the challenging non-enhancing tumor core (NETC), Dice increased by 5.3 %. Optimal regularization weights depended on tissue type and lesion size: λ = 1.0 favored extensive edema and whole-tumor regions, λ = 0.01 best served large contrast-enhancing tumors and necrotic cores. Scenario-weighted evaluation revealed context-dependent optimal models: PI-DynUNet (λ = 0.01) excelled in enhancing-weighted scenarios (RANO: +2.6 %; RT-GTV: +2.2 % vs. baseline), while λ = 1.0 demonstrated superior balanced accuracy (RT-CTV: +1.8 %; Surgical: +1.6 %). CONCLUSIONS: Physics-informed deep learning provides modest but measurable gains in brain metastasis segmentation, and these gains transfer across institutions: external validation on the Stanford BrainMetShare cohort (N = 105) showed that five of seven regularization weights significantly outperform the DynUNet baseline on tumor-core Dice (paired Wilcoxon p < 0.05), with the largest improvement of +10.1 % (p < 0.001) at λ = 0.001 and a ∼6× reduction in inter-fold variance at λ = 1.0. Optimal configuration varies by clinical application, informing context-specific deployment.
Pulmonary nodule growth prediction with anisotropic reaction-diffusion.
Cai R, Zhao H, Yan Y … +3 more , He K, Yan J, Liu B
Comput Methods Programs Biomed · 2026 Sep · PMID 42167009 · Publisher ↗
BACKGROUND AND OBJECTIVE: Accurate prediction of pulmonary nodule growth is critical for early malignancy assessment and timely lung cancer diagnosis. However, pulmonary nodule growth is a complex biological process infl...
BACKGROUND AND OBJECTIVE: Accurate prediction of pulmonary nodule growth is critical for early malignancy assessment and timely lung cancer diagnosis. However, pulmonary nodule growth is a complex biological process influenced by factors such as cellular proliferation, nutrient diffusion, and tissue microenvironment, which are usually nodule-specific and overlooked by traditional prediction models. METHODS: Specifically, we leverage a reaction-diffusion system to exploit properties of a nodule and its surrounding parenchyma for predicting its growth trend, and implement the system using a convolutional operation. To achieve specific information about nodules, we employ a vision transformer to estimate the parameters of the reaction-diffusion system based on consecutive computed tomography scans of a nodule. By doing this, we integrate the reaction-diffusion mathematical modeling with deep neural networks to accurately predict future morphology of pulmonary nodules. RESULTS: We conduct experiments on the benchmark dataset from the National Lung Screening Trial (NLST) to demonstrate the effectiveness of our method. In addition, we also evaluate our method on an in-house dataset to validate its generalization ability and practicality. In particular, RD-ViT reduces volume and mass growth-prediction errors by approximately 50%-80%. CONCLUSIONS: Our method extracts nodule-specific information to accurately forecast future morphologies, validated on benchmark and in-house datasets. It offers a promising tool for personalized lung nodule management, enabling optimized surveillance and enhanced early detection of malignant transformation.
The latent shape space of intracranial saccular aneurysms.
Eulzer P, Voigt H, Meuschke M … +1 more , Lawonn K
Comput Methods Programs Biomed · 2026 Aug · PMID 42161036 · Publisher ↗
BACKGROUND AND OBJECTIVE: Underlying biomechanical instability of the vessel wall is believed to drive the substantial morphological variability observed in saccular intracranial aneurysms. Existing approaches to quantif...
BACKGROUND AND OBJECTIVE: Underlying biomechanical instability of the vessel wall is believed to drive the substantial morphological variability observed in saccular intracranial aneurysms. Existing approaches to quantify this shape variance rely largely on handcrafted descriptors, which capture only limited geometric complexity and show inconsistent findings across studies. Deep learning-based models have shown promise for rupture risk classification but do not provide explicit, analyzable representations of shape. We therefore develop a unified framework that learns a compact but expressive latent representation of aneurysm morphology for generative modeling and rupture-label classification. METHODS: Dense point correspondences were computed for 958 patient-derived aneurysm surfaces (338 ruptured) from five public datasets using uniform parametric mapping. Autoencoder and variational autoencoder models were trained on corresponded meshes at three resolutions (700, 3k, and 12k points) to learn 2-dimensional latent spaces. Reconstruction was evaluated by mean squared error, volumetric error, and Hausdorff distance. We tested rupture-label classification from latent features using logistic regression, support vector machines, and k-nearest neighbors. Baselines included statistical shape models, diffusion maps, UMAP, and established morphological descriptors. RESULTS: The 2-dimensional variational autoencoder achieved high-fidelity reconstruction (mean Hausdorff distance 0.27 ± 0.24 mm), comparable to a 50-dimensional principal component model. Latent features outperformed handcrafted descriptors and other dimensionality-reduction baselines for rupture-label classification, reaching AUC 0.78 and accuracy 0.76 across classifiers (only shape/size parameters). The learned latent spaces showed interpretable continuous transitions between morphological phenotypes, including elongation and multilobularity. CONCLUSIONS: The proposed framework unifies correspondence mapping, generative modeling, and discriminative analysis in a single workflow. The 2-dimensional latent space preserves clinically relevant geometry, enables real-time synthesis and exploration of aneurysm variability, and improves rupture-label discrimination over existing techniques, providing a scalable and interpretable basis for quantitative aneurysm morphometry.
Predicting Avatrombopag response in children with immune thrombocytopenia: A multi-view learning framework for tabular missing data.
Wang Y, Tang Y, Cheng X … +4 more , Lin X, Chen Z, Wu R, Zhang W
Comput Methods Programs Biomed · 2026 Aug · PMID 42161035 · Publisher ↗
BACKGROUND AND OBJECTIVE: Avatrombopag is a thrombopoietin receptor agonist that has demonstrated clinical efficacy in rapidly increasing platelet counts and achieving bleeding control in children with chronic immune thr...
BACKGROUND AND OBJECTIVE: Avatrombopag is a thrombopoietin receptor agonist that has demonstrated clinical efficacy in rapidly increasing platelet counts and achieving bleeding control in children with chronic immune thrombocytopenia. Early prediction of treatment response is critical for reducing unnecessary drug exposure and guiding personalized treatment strategies. However, accurate prediction remains challenging due to the inherent characteristics of clinical tabular data, including missing values caused by inconsistent data recording, small sample size, and high-dimensional clinical features. This study aims to develop an effective computational framework for early prediction of treatment response under these conditions. METHODS: We propose a multi-view learning framework that constructs multiple views by applying diverse imputation methods to the same dataset, with each view capturing unique assumptions about missingness. A feature selection encoder is employed to reduce feature redundancy and improve model interpretability. Multi-view fusion and co-regularization are integrated at the prediction level to learn complementary patterns across different views. In addition, contrastive learning is introduced to alleviate the small data problem and enhance representation robustness. RESULTS: Experiments on real-world clinical data of children treated with avatrombopag demonstrate that the proposed method consistently outperforms multiple competitive baseline models in predicting treatment response. CONCLUSIONS: This study provides a practical and interpretable multi-view learning framework for early identification of treatment response in chronic immune thrombocytopenia, supporting more personalized and efficient treatment decisions in pediatric hematology.
Aortic stiffening after thoracic aortic stent grafting: A multi-patient specific computational study.
Molinari L, Cebull H, Piccinelli M … +2 more , Leshnower BG, Veneziani A
Comput Methods Programs Biomed · 2026 Aug · PMID 42161034 · Publisher ↗
BACKGROUND AND OBJECTIVE: Treatment of Descending Thoracic Aortic Aneurysms (DTAA) using open surgical or endovascular methods (Thoracic EndoVascular Aortic Repair) is widely accepted in medicine. However, these procedur...
BACKGROUND AND OBJECTIVE: Treatment of Descending Thoracic Aortic Aneurysms (DTAA) using open surgical or endovascular methods (Thoracic EndoVascular Aortic Repair) is widely accepted in medicine. However, these procedures modify the aorta's anatomy and biomechanics, triggering anomalous wave reflections and cardiac remodeling. The complex interplay among these factors is largely unexplored, hampering procedural efficacy and long-term predictability. Computational fluid structure interaction (FSI) is a powerful tool for exploring these dynamics, but the computational complexity of simulating 3D scenarios often poses practical challenges, in terms of efficiency and reliability. To address this, we employ cost-effective geometrical multiscale 0D-1D FSI models. METHODS: We integrate a simplified lumped parameter model of the left heart and an extended 1D systemic circulation model (covering the largest 55 arteries), implemented in the Multiscale module in the C++ finite element library LifeV. Patient-specific preoperative and postoperative 1D-FSI models were derived from CT angiography data of 11 patients (6 open surgery, 5 TEVAR), with implant stiffness adhering to established literature values (1.2 MPa Dacron grafts, 51.7 MPa metallic stents). Physiological inflow conditions are imposed in the ascending aorta, while three-element Windkessel models account for peripheral circulation. We simulated a total of 22 cases (baseline and postoperative for each case, spanning five heartbeats each) and compared the peak systolic pressure, pressure-volume loops (PV) and pulse-wave velocity (PWV) between vessels. RESULTS: Our findings demonstrate that the presence of implants can increase ascending aorta pressure due to the backpropagation of pressure waves. The cases of TEVAR exhibit significantly higher PWV. Complex geometric cases exhibit smoother pressure-flow profiles after surgery. In general, the combination of altered arterial geometry and stiffness in postoperative conditions can significantly alter cardiac PV loops. We quantitatively demonstrate that descending aortic elongation independently compensates for material stiffness effects, with a strong negative correlation (r=-0.762,p=0.0065) between length change and post-operative pressure elevation. Cases achieving ≥15% descending elongation uniformly showed pressure reduction or minimal elevation. CONCLUSIONS: Extending these insights to larger cohorts of patients has the potential to reveal mechanisms shaping the long-term effects of DTAA repair. Our results highlight the crucial need for a combined analysis of both stiffness and geometrical changes following surgery, as a more significant local stiffening (such as that caused by a stent) does not necessarily lead to cardiac overload if simultaneous changes in aortic geometry can compensate for the stent's impact.
Zero-shot arbitrary-scale super resolution in susceptibility-weighted imaging for cerebral microbleed analysis.
Liu F, Zhang R, Lv Z … +4 more , Zhao J, Wu X, Xie G, Guo L
Comput Methods Programs Biomed · 2026 Aug · PMID 42155350 · Publisher ↗
BACKGROUND AND OBJECTIVE: Susceptibility Weighted Imaging (SWI) plays a pivotal role in detecting cerebral microbleeds (CMBs), key biomarkers of vascular abnormalities and neurodegenerative diseases. Clinical protocols o...
BACKGROUND AND OBJECTIVE: Susceptibility Weighted Imaging (SWI) plays a pivotal role in detecting cerebral microbleeds (CMBs), key biomarkers of vascular abnormalities and neurodegenerative diseases. Clinical protocols often use larger slice spacing to reduce scan time, resulting in low-resolution SWI with compromised through-plane detail. Although deep learning-based super-resolution (SR) methods show promise, they require large paired high-resolution (HR) and low-resolution (LR) datasets that are difficult to acquire in medical imaging. To address these challenges, we propose MagNeRF, a zero-shot, single-subject arbitrary-scale SR framework that learns an implicit prior from a single LR volume without external paired training data. METHODS: MagNeRF introduces three key innovations for SWI: (1) a dilated patch-based sampling strategy to improve spatial context and local detail recovery; (2) a spherical sampling strategy to capture the radial gradient decay of magnetic susceptibility signals in SWI; and (3) adaptive loss functions (adaptive multi-scale structural similarity and adaptive mean squared error) that emphasize perceptual fidelity and structural preservation. RESULTS: We rigorously evaluated MagNeRF on two SWI datasets targeting CMBs. Results demonstrate that MagNeRF outperforms state-of-the-art methods, producing reconstructed SWI volumes with high visual fidelity and preserved diagnostically relevant structures. Further validation on a T1-weighted dataset and a real-world LR T2*-weighted dataset confirms its robustness across diverse MRI contrasts. Notably, downstream CMB lesion segmentation using the reconstructed HR images achieves performance closely approaching that of the original HR data, underscoring the clinical utility of the proposed approach. CONCLUSIONS: MagNeRF shows significant potential in preserving clinically meaningful microbleed features in SWI. By enabling HR reconstruction from LR inputs, MagNeRF reduces patient burden, enhances diagnostic accuracy, and broadens the clinical applicability of SWI.
Detecting cardiovascular diseases using ECG scans and explainable artificial intelligence.
Czerwinski A, Kucharski D, Kawa J … +4 more , Zheng Y, Lip GYH, Nalepa J, Wijata AM
Comput Methods Programs Biomed · 2026 Aug · PMID 42155349 · Publisher ↗
BACKGROUND AND OBJECTIVE: Cardiovascular diseases are the leading cause of mortality globally, requiring early and accurate detection through tools like electrocardiography. While artificial intelligence models have emer...
BACKGROUND AND OBJECTIVE: Cardiovascular diseases are the leading cause of mortality globally, requiring early and accurate detection through tools like electrocardiography. While artificial intelligence models have emerged to provide reproducible analysis of electrocardiogram printouts, their clinical deployment is hindered by a lack of transparency and sensitivity to real-world image variations such as discolorations, handwriting, or paper wrinkles. This study introduces an explainable artificial intelligence framework designed to quantify the stability of deep learning models and identify vulnerabilities in their behavior under controlled image perturbations. METHODS: We utilized a large-scale dataset of electrocardiogram printouts synthesized from the PTB-XL benchmark, creating both clean and contaminated versions featuring various image-level manipulations. Four deep learning architectures, including EfficientNet and InceptionNet, were trained and evaluated using different activation functions. The stability of these models was assessed using local interpretable model-agnostic explanations. We employed intersection over union metrics to measure the consistency of explanations across perturbations and extracted radiomic-like image features to quantitatively analyze the characteristics of the generated explanations. RESULTS: Our experiments demonstrate that models trained on augmented datasets generalize better to perturbed data, with the best-performing model achieving an area under the receiver operating characteristic curve of 0.894 on the contaminated test set. Stability analysis showed that models trained on data containing perturbations achieved the highest average intersection over union of 0.399, indicating a more consistent focus on diagnostic features. Furthermore, radiomic-like features enabled the precise identification of the underlying deep learning model with an accuracy of up to 98%. CONCLUSIONS: The proposed framework enables a comprehensive visual and quantitative evaluation of artificial intelligence stability in cardiovascular disease detection. By identifying how specific image manipulations affect model reliability, this approach can guide the development of more robust algorithms and targeted data augmentation strategies. To ensure full reproducibility and foster cross-domain collaboration, our tools and datasets are available at https://github.com/smile-research/xai-ecg.
A 3D parametric model of the endothelial monolayer to predict its barrier integrity: Influence of mechanical and geometric conditions.
Paseta O, García-Aznar JM, Gómez-Benito MJ
Comput Methods Programs Biomed · 2026 Aug · PMID 42143961 · Publisher ↗
BACKGROUND AND OBJECTIVE: The endothelium, which consists on a monolayer of endothelial cells, lines the interior surface of blood vessels. These cells form a dynamic structure in which adhesions between them are constan...
BACKGROUND AND OBJECTIVE: The endothelium, which consists on a monolayer of endothelial cells, lines the interior surface of blood vessels. These cells form a dynamic structure in which adhesions between them are constantly breaking down and rebuilding. This process creates gaps in the endothelium through which various substances and different cell types (including cancer cells) can intra/extravasate. It is known that mechanical and geometrical factors, such as the stiffness of the extracellular matrix surrounding the endothelial monolayer and its curvature, influence gap formation in the endothelium, thereby regulating its barrier integrity. This study uses finite element simulation based on a parametric model to investigate the regulatory role of these factors on gap opening in the endothelial monolayer. METHODS: We developed a parametric geometry model that enables the rapid generation of an endothelial monolayer and its surrounding extracellular matrix. We studied the intercellular stresses and gap formation on the monolayer for different matrix stiffnesses and vessel diameters using this geometry and finite element simulations. RESULTS: Our simulations indicate that stiffer extracellular matrices are associated with higher intercellular stresses and a greater percentage of gaps in the monolayer, regardless of substrate shape or vessel diameter. At a constant substrate stiffness, both intercellular stresses and gap formation increase as the vessel diameter decreases. CONCLUSIONS: Our findings suggest that matrix stiffness and vessel curvature synergistically impair endothelial barrier integrity. Specifically, the combination of high substrate stiffness and reduced vessel diameter significantly increases the susceptibility to monolayer gap formation.
The challenge of data scarcity and imbalanced classes in radiomics performance.
Rodriguez-Belenguer P, Beser-Robles M, Marfil-Trujillo M … +2 more , Cerdá-Alberich L, Martí-Bonmatí L
Comput Methods Programs Biomed · 2026 Aug · PMID 42143960 · Publisher ↗
BACKGROUND AND OBJECTIVES: Radiomics-based machine learning models are increasingly used for clinical decision-making, yet their reliability is often constrained by limited sample sizes and class imbalance in imaging dat...
BACKGROUND AND OBJECTIVES: Radiomics-based machine learning models are increasingly used for clinical decision-making, yet their reliability is often constrained by limited sample sizes and class imbalance in imaging datasets. While prior studies addressed these challenges independently, their combined impact remains insufficiently explored. This study quantifies the individual and joint effects of progressive data scarcity and class imbalance on radiomics model performance, aiming to identify mitigation strategies. METHODS: Three radiomics datasets-PI-CAI (clinically significant prostate cancer detection), BraTS2021 (MGMT promoter methylation prediction in glioblastoma), and Hunter2023 (lung nodule malignancy classification)-were analyzed under four conditions: baseline (balanced, fixed-size dataset), progressive class imbalance, progressive sample size reduction, and a combined scenario. Five machine learning models were evaluated, with XGB selected for PI-CAI and Hunter2023 and MLP for BraTS2021 based on highest balanced accuracy. Imbalance was addressed using state-of-the-art sampling techniques, and data scarcity was mitigated using Tabular Variational Autoencoders (TVAE). Statistical significance was evaluated using paired permutation t-tests across repeated cross-validation splits. RESULTS: Feature selection played a key role in model performance and interpretability. The most predictive features were biologically plausible and dataset-specific, such as perinodular texture heterogeneity in lung cancer or gray-level non-uniformity in glioblastoma. Under progressive class imbalance, unsampled models showed degradation in balanced accuracy. Applying the best-performing sampling strategy improved balanced accuracy in PI-CAI (Δ = 0.065; dz = 1.54), BraTS2021 (Δ = 0.057; dz = 1.25), and Hunter2023 (Δ = 0.087; dz = 1.35), all p < 0.001. Under progressive sample size reduction, TVAE showed dataset-dependent effects: no significant change in PI-CAI, moderate degradation in BraTS2021 (Δ = -0.042; dz = -0.60; p < 0.001), and improvement in Hunter2023 (Δ = 0.029; dz = 0.46; p < 0.001). In the combined scenario, TVAE plus sampling increased balanced accuracy in PI-CAI (Δ = 0.023; dz = 0.42; p = 0.03), BraTS2021 (Δ = 0.070; dz = 0.95; p < 0.001), and Hunter2023 (Δ = 0.041; dz = 0.61; p < 0.01). CONCLUSION: Class imbalance consistently impairs radiomics performance and may be mitigated through undersampling-based strategies. The impact of augmentation under data scarcity is dataset-dependent, but combined correction strategies yield gains under compounded constraints.
Glo-MMF: A modular multi-model framework for automated morphometry of glomerular ultrastructural features.
Zhang Z, Weng D, Zhang G … +7 more , Chen X, Long K, Geng J, Lu Y, Zhang L, Zhou Z, Cao L
Comput Methods Programs Biomed · 2026 Aug · PMID 42139796 · Publisher ↗
BACKGROUND AND OBJECTIVE: Automated morphological analysis of glomerular ultrastructures facilitates diagnosis by reducing pathologists' burden and improving efficiency and accuracy. However, the inherent heterogeneity o...
BACKGROUND AND OBJECTIVE: Automated morphological analysis of glomerular ultrastructures facilitates diagnosis by reducing pathologists' burden and improving efficiency and accuracy. However, the inherent heterogeneity of different ultrastructural quantification processes hinders the ability of existing single-model studies to fulfill the clinical demand for the simultaneous achievement of multiple analytical objectives, including structure measurement, state assessment, and lesion localization. To address this, we developed Glo-MMF, a modular framework that integrates automated analysis, incorporating segmentation, classification, and detection. This framework aims to jointly quantify key ultrastructural features within the glomerulus, offering strong support for renal pathology research and diagnostic assistance. METHODS: Glo-MMF decomposes the quantification task of glomerular ultrastructural morphological features by constructing three dedicated deep learning models: an ultrastructural segmentation model, a glomerular filtration barrier (GFB) region classification model, and an electron-dense deposits (EDD) detection model. The outputs of these models are systematically integrated through a post-processing workflow comprising four computer vision modules, enabling the measurement of multiple ultrastructural features. Key operations include adaptive cropping of GFB regions and screening of suitable measurement locations. This approach significantly enhances measurement reliability, overcomes the limitations of traditional grading descriptions, and provides more thorough and interpretable quantitative results for glomerular pathological analysis. RESULTS: Trained on 372 renal biopsy electron microscopy images, the Glo-MMF framework enables simultaneous quantification of the thickness of glomerular basement membrane (GBM), the degree of foot process effacement (FPE), and the location of EDD. In 115 test cases spanning 9 renal pathological types, the automated quantification results for these three features demonstrated strong agreement with descriptions in pathological reports. Processing and analysis per case in a CPU environment, including measurement of GBM thickness, quantification of FPE degree, and location of EDD, required an average time of 4.23 ± 0.48 s. CONCLUSIONS: The modular design of Glo-MMF enables a certain degree of flexible extensibility, supporting the joint quantification of multiple key glomerular ultrastructural features. This framework ensures robust performance and clinical applicability across various renal pathological types, demonstrating significant potential to play an efficient auxiliary role in glomerular pathological analysis.
Antiangiogenic therapy enhances CAR-T cell efficacy in solid tumors: Insights from a hybrid multiscale model.
Mortazavi SMA, Firoozabadi B
Comput Methods Programs Biomed · 2026 Aug · PMID 42139795 · Publisher ↗
BACKGROUND AND OBJECTIVES: Despite the remarkable success of chimeric antigen receptor (CAR) T cells in hematological malignancies, their therapeutic efficacy in solid tumors is still limited, largely due to the insuffic...
BACKGROUND AND OBJECTIVES: Despite the remarkable success of chimeric antigen receptor (CAR) T cells in hematological malignancies, their therapeutic efficacy in solid tumors is still limited, largely due to the insufficient infiltration of CAR-T cells and the immunosuppressive tumor microenvironment (TME). Recently, murine studies have shown that anti-angiogenic therapy can improve the efficacy of CAR-T cells. However, extensive research is required to unveil the underlying mechanism and develop effective therapeutic protocols. METHODS: Here, we developed a hybrid discrete - continuous multi-scale computational model to study solid tumor growth, angiogenesis, and CAR-T cell infiltration. Our model considers the intravenous infusion of CAR-T cells, their extravasation through the endothelium, and their cytotoxic effect in the TME. Therefore, this model enables us to robustly investigate the impact of anti-angiogenic therapy on the efficacy of CAR-T cells. RESULTS: Our findings showed that anti-angiogenic therapy enhances the delivery and efficacy of CAR-T cells in solid tumors. Subsequently, we evaluated different therapeutic regimens. The results revealed that although both adjuvant and neoadjuvant anti-angiogenic therapy can improve outcome, the neoadjuvant regimen yields higher survival rates. Moreover, our model suggested an optimal dosing schedule to maximize survival. CONCLUSIONS: Our results provide profound insight into the therapeutic dynamics of combined anti-angiogenic and CAR-T cell therapy. Additionally, the model establishes a computational framework for selecting treatment regimens for clinical trials.
Morphology-, noise-, and resolution-robust ultrasound elasticity imaging with Fourier neural operator.
Kim H, Lee H, Park M … +1 more , Ryu S
Comput Methods Programs Biomed · 2026 Aug · PMID 42134076 · Publisher ↗
Quasi-static strain elastography is a non-invasive technique for estimating tissue stiffness fields from displacement fields obtained by comparing ultrasound signals before and after compression. While recent deep learni...
Quasi-static strain elastography is a non-invasive technique for estimating tissue stiffness fields from displacement fields obtained by comparing ultrasound signals before and after compression. While recent deep learning approaches have enabled faster and more accurate elasticity estimation compared to traditional methods, several challenges remain for clinical translation. In this study, we employed finite element simulations of free-hand palpation to investigate the applicability of the Fourier neural operator (FNO) for mapping displacement fields to elasticity fields in quasi-static strain elastography. Four practical scenarios were investigated: (1) prediction across diverse lesion morphologies, (2) generalization to cases with lesion counts differing from those in the training data and to lesion morphologies not seen during training, including adaptation through few-shot fine-tuning, (3) robustness to noise in measured displacement fields, and (4) resilience to variations in ultrasound device resolution. FNO achieved competitive predictive accuracy across lesion types and consistently outperformed DeepONet, although U-Net yielded lower errors in several cases involving multiple lesions or sharp modulus discontinuities. In contrast, FNO showed markedly stronger robustness to noisy displacement inputs than U-Net and maintained reliable performance under moderate resolution mismatch, with degradation observed under more aggressive downsampling. For unseen realistic tumor morphologies, few-shot fine-tuning substantially improved prediction accuracy. Finally, we performed a qualitative zero-shot evaluation on a public experimental phantom dataset using axial displacement fields estimated from pre- and post-compression radio-frequency (RF) data. The simulation-trained FNO localized inclusion-like regions without additional fine-tuning, although quantitative contrast agreement remained limited because pixelwise modulus ground truth was unavailable. These results suggest that FNO is a promising operator-learning framework for axial displacement-to-modulus mapping in quasi-static strain elastography, while explicit simulation of raw ultrasound signal formation and displacement estimation from pre- and post-compression RF/in-phase quadrature (IQ) data remains outside the scope of the present study.
High-resolution 3D flow reconstruction of cerebrospinal fluid microcirculation using physics-informed neural network: Conceptualization and application to a large animal model.
Epshtein M, Mekler T, Shazeeb MS … +4 more , Lindsay C, Gounis MJ, Korin N, Anagnostakou V
Comput Methods Programs Biomed · 2026 Aug · PMID 42114465 · Publisher ↗
In this study, we present a novel Physics-Informed Neural Network (PINN) framework that reconstructs 3D flows using planar velocity projections from arbitrarily oriented planes. The method is designed for the reconstruct...
In this study, we present a novel Physics-Informed Neural Network (PINN) framework that reconstructs 3D flows using planar velocity projections from arbitrarily oriented planes. The method is designed for the reconstruction of low-Reynolds-number flows typical of cerebrospinal fluid (CSF) in the subarachnoid space (SAS). The method utilizes a projection loss function combined with gradient smoothing regularization during network training. We show that plane orientation with perturbations of 0.05 (relative to the main flow axis) or greater is sufficient axial data for accurate reconstruction. Additionally, gradient exponential moving average smoothing with amplification improves convergence and stability, particularly for near-parallel planes of acquisition. The method was compared against computational fluid dynamics (CFD) data and applied to flow in a realistic canine SAS geometry derived from intravascular optical coherence tomography (OCT), demonstrating the framework's potential for in-vivo CSF flow imaging.
Uncertainty‑aware sepsis survival prediction using conformal XGBoost on minimal clinical features under Sepsis‑3 criteria.
Parreño SJ
Comput Methods Programs Biomed · 2026 Aug · PMID 42107321 · Publisher ↗
BACKGROUND AND OBJECTIVE: Sepsis remains a major cause of in-hospital mortality, but many prognostic models require extensive clinical data. This study evaluated whether a minimal-feature XGBoost model combined with indu...
BACKGROUND AND OBJECTIVE: Sepsis remains a major cause of in-hospital mortality, but many prognostic models require extensive clinical data. This study evaluated whether a minimal-feature XGBoost model combined with inductive conformal prediction could support sepsis survival prediction and uncertainty quantification under Sepsis-3 criteria. METHODS: We performed a retrospective secondary analysis of 19,051 Norwegian Sepsis-3 admissions and externally validated the model in 137 South Korean cases. Predictors were age, sex, and septic episode number. The Norwegian cohort was split into proper-training, calibration, and internal test sets. Class imbalance in the proper-training split was handled using Random Over-Sampling Examples (ROSE), and XGBoost hyperparameters were tuned by grid search with 5-fold stratified cross-validation. Model evaluation included confusion matrices, ROC AUC, PR AUC, calibration analysis, and inductive conformal prediction at 80 %, 90 %, and 95 % nominal coverage. Ninety-five percent confidence intervals were estimated by stratified bootstrap resampling. RESULTS: All required variables were complete in both cohorts. At the default threshold, the model classified all cases as survival in both cohorts, yielding accuracies of 0.811 and 0.825 but specificity, Cohen's kappa, and Matthews correlation coefficient of 0.000. ROC AUC was 0.578 (95 % CI 0.556-0.601) internally and 0.570 (95 % CI 0.442-0.690) externally. At 90 % nominal coverage, conformal prediction achieved empirical coverages of 0.901 (95 % CI 0.894-0.908) internally and 0.876 (95 % CI 0.847-0.912) externally, with mean prediction-set sizes of 1.411 and 1.146. CONCLUSIONS: The conformal wrapper maintained near-target empirical coverage, but the underlying three-variable classifier showed limited discrimination. Additional predictors and broader validation are needed before clinical use.
From 2D to 3D: Automated ultrasound segmentation and cross-sectional validation in murine tumor models.
Smolak-Dyżewska W, Bazak J, Brandys W … +10 more , Bienia A, Murzyn A, Płóciennik B, Drwięga G, Kozik J, Drzał A, Leszczyński B, Spurek P, Elas M, Krzykawska-Serda M
Comput Methods Programs Biomed · 2026 Aug · PMID 42107155 · Publisher ↗
BACKGROUND AND OBJECTIVE: Ultrasound (US) is a widely used method for non-invasive tumor monitoring. Semantic segmentation of tumors in US imagery is a necessary step to reconstruct 3D geometry of a region of interest (R...
BACKGROUND AND OBJECTIVE: Ultrasound (US) is a widely used method for non-invasive tumor monitoring. Semantic segmentation of tumors in US imagery is a necessary step to reconstruct 3D geometry of a region of interest (ROI). Still, the segmentation task remains challenging due to variable signal-to-noise ratio (SNR) and modest soft-tissue contrast in US imaging. Our objective was to generate a new dataset of murine tumors and present a standardized pipeline for 3D volume reconstruction, suitable for preclinical research. METHODS: Human LN229 and murine 4T1, PanO2, B16 and LLC tumors were imaged in vivo using high-frequency US systems (Vevo F2, Vevo 2100). Our dataset comprised 3442 images, including expert-curated masks and a challenging out-of-distribution (OOD) test set. We evaluated U-Net, Res U-Net, Attention U-Net, and R2AU-Net, with and without autoencoder pretraining on unlabeled frames. For volumetry, 2D masks were converted to voxel grids using known imaging geometry; through-plane interpolation and Marching Cubes surface extraction enabled shape-agnostic 3D volume computation. We compared US-derived volumes with micro-CT, calipers, mold-based, and tumor weight. RESULTS: Across random train-test splits, all models achieved Dice > 0.90. On the subject-independent special testing dataset, performance decreased, indicating limited generalization under distribution shift; the pretrained Attention Res U-Net achieved the highest overlap (Dice 0.750, IoU 0.604), while the pretrained Attention U-Net also remained comparatively robust (Dice 0.731). The 3D reconstruction pipeline produced consistent longitudinal volumes, and cross-modal comparison showed that US-based volumes agreed with micro-CT and physical measurements. CONCLUSIONS: This study presents a standardized workflow for automated tumor segmentation and 3D ultrasound-based volumetry, enabling more objective and reproducible assessment of tumor burden in preclinical oncology studies.
A novel qVGG-4 model for optimizing a parameterized quantum circuit in a quantum-IoT-based brain tumor detection and monitoring system.
Ahad MT, Song B, Li Y
Comput Methods Programs Biomed · 2026 Aug · PMID 42105676 · Publisher ↗
BACKGROUND: Since brain tumors (BTs) require early detection for timely and effective treatment planning, this study presents two quantum deep learning (Q-DL) approaches: a quantum Convolutional Neural Network (CNN) and...
BACKGROUND: Since brain tumors (BTs) require early detection for timely and effective treatment planning, this study presents two quantum deep learning (Q-DL) approaches: a quantum Convolutional Neural Network (CNN) and a quantum Vision Transformer (ViT). The implications of Q-DL for disease detection in medical images are limited, and previous studies have suggested that Q-DL has unsatisfactory accuracy. METHODS: To fill this gap, four models, (1) quantum CNN (Q-CNN), (2) hybrid quantum CNN (HQ-CNN), (3) Q-ViT, and (4) hybrid quantum ViT (HQ-ViT), were developed and tested on four BT-MRI datasets. BT patients demand real-time monitoring as they suffer from headaches, seizures, cognitive and behavioral changes, and neurological deficits. Therefore, we propose a smart brain tumor management system (SBTM) for real-time monitoring. RESULTS: Trained on the three brain tumor datasets using the Adam optimizer and five-fold cross-validation, the hybrid Q-DL, HQ-CNN, achieved an accuracy of 97%, and HQ-ViT achieved 96% in (tumor, no tumor) classification, which outperforms the parameterized quantum circuit (PQC)-based Q-DLs. The high accuracy of hybrid models continues: in 3 classes, 44% by CNN and 28% by HQ-ViT, and in 4 classes, 49% by HQ-CNN and 23% by HQ-ViT. The increased accuracy of hybrid models continues in the detection and classification test dataset of brain tumor MRI images. The results suggest that combining qVGG-4 with PQC in both CNNs and ViTs yields more powerful feature extraction than either alone. CONCLUSIONS: The main novelty of this study is the use of a qVGG-4 model that optimizes PQC. Whereas traditional CNNs struggle with small tumors, the HQ-CNN and HQ-ViT methods achieve impressive accuracy even on 28 × 28-pixel images. This result solves the issue of handling complex lesion detection in small areas and accelerates the model training time. The high accuracy in detecting and classifying unseen MRI images is a significant contribution to SBTM. In clinical settings, a machine learning model is expected to perform well in detecting and classifying new MRI images.
Neural estimates of language comprehension of sentences with selected action verbs in a non-literal context in the Polish language studied by permutation cluster-based analysis and decoding of the event-related potentials.
Maciejewska K, Wałczyk K, Nowak T
Comput Methods Programs Biomed · 2026 Aug · PMID 42105445 · Publisher ↗
BACKGROUND: This work examines how the human brain processes mental metaphors in Polish verbal phraseologisms (a fixed, non-compositional combination of at least two elements that conveys a single meaning distinct from i...
BACKGROUND: This work examines how the human brain processes mental metaphors in Polish verbal phraseologisms (a fixed, non-compositional combination of at least two elements that conveys a single meaning distinct from its parts and functions as a pragmatically expressive, semantically metaphorical, and syntactically unified unit) by studying dynamic changes in brain electrical activity using electroencephalography (EEG). The study aimed to investigate how figurative motor expressions (known and unknown) and known literal motor and mental expressions are processed via event-related potentials (ERPs). METHODS: EEG was recorded from 34 participants using a 32-channel active EEG system in an extended 10-20 montage. We used two advanced biomedical information processing methods: (1) the nonparametric cluster-based permutation analysis, which doesn't require specifying a priori time and spatial boundaries of the anticipated effects, and (2) machine learning (ML) using multivariate pattern analysis (MVPA) to decode the ERPs between the studied conditions. RESULTS: The comparison of the ERPs elicited by short sentences with an action verb in literal and non-literal (phraseological and unknown figurative) contexts, and a mental literal verb as a control condition, presented to participants on the computer screen, showed a decrease in amplitude from phraseological -> unknown figurative -> motor literal -> mental literal. The first effect was a robust and stable difference between the phraseological and literal conditions, corresponding to a spatiotemporal cluster maximal over frontal and central scalp areas, starting around 550 ms after noun phrase onset, as confirmed by machine-learning decoding. This effect likely represents a late frontal positivity (LFP). The second one showed a weaker difference between the non-literal and mental literal conditions, corresponding to a spatiotemporal cluster maximal over the midline central area around 250-350 ms after noun phrase onset. This effect relates to a higher P2-N2 complex in both figurative conditions compared to the mental one. CONCLUSIONS: Our results suggest a specific way in which the human brain processes action verbs in non-literal contexts, thereby filling a gap in neurocognitive research by providing a deeper understanding of how the brain processes language. They may aid neuroengineering research in designing more ecologically valid biomedical engineering solutions, such as more accurate language models or human-robot interactions.
About
- Frequency
- Sun
- Papers found
- 200
- RSS feed
- Subscribe