Computer Methods And Programs In Biomedicine[JOURNAL]
Sun
200 papers
Synaptic regulatory flux dynamics: Stiffness and homeostatic failure in Alzheimer's disease.
Srimaharaj W
Comput Methods Programs Biomed · 2026 Jun · PMID 42308717 · Publisher ↗
BACKGROUND: Traditional electroencephalographic analysis in Alzheimer Disease focuses on spectral slowing, a terminal phenotype that overlooks the preceding functional failure of homeostasis. This study introduces Synapt...
BACKGROUND: Traditional electroencephalographic analysis in Alzheimer Disease focuses on spectral slowing, a terminal phenotype that overlooks the preceding functional failure of homeostasis. This study introduces Synaptic Regulatory Flux Dynamics (SRFD), a computational framework that models the EEG signal as a physical trajectory governed by the interplay between Excitatory Drive (E) and Regulatory Flux (R). METHODS: Resting state EEG recordings from 88 participants (36 CE, 23 Frontotemporal Dementia [FTD], and 29 Cognitively Normal [CN]) were analyzed. The study derived novel biophysical metrics, including Instantaneous Homeostatic Error (IHE) and the Synaptic Stiffness Index (SSI), to assess the integrity of the excitation inhibition (E/I) balance. RESULTS: It found that the AD cohort exhibited a statistically significant elevation in Mean IHE compared to controls (0.557vs0.535; p=0.0115), indicating a systemic failure of inhibitory interneurons to clamp excitatory transients. This flux decoupling was accompanied by a reduced Synaptic Stiffness Index (p=0.011), predominantly affecting posterior beta band networks. Furthermore, the framework successfully differentiated the metabolic silence of AD from the high energy, disorganized hyperexcitability of FTD. CONCLUSION: Synaptic Regulatory Flux Dynamics provides a noninvasive, mechanistic biomarker of synaptic fragility that precedes gross atrophy. By evaluating the loss of regulatory stiffness, this framework offers a functional window into the early pathophysiology of neurodegeneration, potentially enhancing early detection and therapeutic monitoring.
BoneContourNet-Vis: a lightweight end-to-end deep learning framework for real-time ultrasound bone imaging in orthopedic surgery.
Wang X, Song L, Qiu H … +6 more , Qin Y, Pan Z, Zhang Z, Zheng Y, Xing L, Yin X
Comput Methods Programs Biomed · 2026 Jun · PMID 42308716 · Publisher ↗
BACKGROUND AND OBJECTIVE: Real-time ultrasound bone visualization is crucial for orthopedic surgical navigation, but current methods rely on a two-stage "segmentation-then-rendering" pipeline that adds latency and separa...
BACKGROUND AND OBJECTIVE: Real-time ultrasound bone visualization is crucial for orthopedic surgical navigation, but current methods rely on a two-stage "segmentation-then-rendering" pipeline that adds latency and separates the enhancement from the original image context. These methods also struggle with discontinuous bone contours due to acoustic shadowing and angle sensitivity, limiting their clinical utility. We aimed to develop a new end-to-end deep learning framework to overcome these challenges and provide real-time, continuous bone visualization. METHODS: We reformulated ultrasound bone visualization as a direct guided overlay task: a deep network predicts a calibrated bone probability map and fuses it with the ultrasound image, eliminating separate mask rendering. Based on this concept, we developed BoneContourNet-Vis, a lightweight end-to-end model built on a ConvNeXt V2 Nano backbone for strong feature extraction and high-throughput inference. The network incorporates three specialized modules: an Edge Attention Module (EAM) to enhance thin cortical edges, a Physics-aware Interaction Module (PIM) to inject acoustic shadow and phase priors into deep features, and a Contour-Adaptive Module (CAM) to enforce smooth, continuous bone contours via graph-based refinement. RESULTS: Comprehensive evaluation on the Bone100K dataset (∼100k ultrasound frames) demonstrated that our method outperforms representative approaches (e.g., HiFormer, MedNeXt, MedSAM) with Dice 0.933, IoU 0.875, Precision 0.96, Recall 0.91, Specificity 0.95, HD95 6.52 px, ASSD 2.07 px, while achieving ∼109 frames per second (9.18 ms latency) - meeting intraoperative real-time requirements. Our approach showed more complete and continuous bone contours under heavy acoustic shadowing and maintained the grayscale context of the original ultrasound. CONCLUSIONS: The proposed BoneContourNet-Vis framework improves the completeness and interpretability of ultrasound bone imaging without sacrificing inference speed. It delivers an accurate and real-time bone visualization solution. In future work, this framework can be extended to multi-planar or dynamic ultrasound sequences to achieve real-time three-dimensional bone reconstruction. Combined with optical tracking and probe calibration, the method can further enhance millimeter-level localization accuracy, thereby providing robust technical support for clinical orthopedic navigation, intraoperative guidance, and rapid bedside fracture assessment.
Validation of the "protective flow attenuation" hypothesis: Numerical and clinical insights into helical flow and thrombotic risk in cirrhosis.
Xiong Z, Xu R, Shi C … +2 more , Zhao J, Zheng T
Comput Methods Programs Biomed · 2026 Jun · PMID 42308715 · Publisher ↗
BACKGROUND AND OBJECTIVES: Portal vein thrombosis (PVT) is a common complication in cirrhosis patients with portal hypertension and is associated with unfavorable clinical outcomes. However, reliable methods for early id...
BACKGROUND AND OBJECTIVES: Portal vein thrombosis (PVT) is a common complication in cirrhosis patients with portal hypertension and is associated with unfavorable clinical outcomes. However, reliable methods for early identification of patients at high risk of PVT remain lacking. We propose the protective flow attenuation hypothesis, which postulates that attenuation of helical flow within the portal vein (PV) promotes thrombus formation. This primary objective of this study was to evaluate this hypothesis through computational fluid dynamics simulations and patient-specific analyses, while the secondarily objective was to assess the potential of Helical intensity (H) as a biomarker for PVT risk stratification. METHODS: A total of 80 patients with liver cirrhosis were retrospectively enrolled, including 40 patients with PVT and 40 without. An ideal PV model was constructed, and a series of parametric models with different spleno-mesenteric confluence (SMC) angles and portal blood flow velocities (PBFV) were generated to investigate the influence of PV geometry and hemodynamic conditions on helical flow characteristics. Patient-specific imaging and hemodynamic data were further analyzed to validate the simulation findings. RESULTS: Smaller SMC angles and lower PBFV were associated with significant attenuation of PV helical flow. Lower Helicity was correlated with a greater proportion of PV wall area exposed to low WSS. Compared with patients without PVT, patients with PVT exhibited significantly smaller SMC angles and lower PV H. H showed superior discrimination for identifying PVT compared with either SMC angle or PBFV alone. In patients who subsequently developed PVT, PV H was higher after thrombus formation than before. In addition, helical thrombus morphology was observed in a subset of patients. CONCLUSIONS: By integrating computational modeling with clinical analysis, this study provides evidence supporting the protective flow attenuation hypothesis, indicating that diminished PV H may contribute to thrombogenesis in cirrhosis. Furthermore, H outperformed conventional geometric and hemodynamic parameters in PVT risk stratification, highlighting its potential as a quantitative biomarker for future prospective validation and clinical translation.
A shortest-path and ADMM-based fluence-level optimization framework for discretized non-coplanar VMAT.
Wang F, Wang Y, Nong F … +5 more , Li C, Li Y, Xu D, Zhuge C, Yang R
Comput Methods Programs Biomed · 2026 Jun · PMID 42308714 · Publisher ↗
BACKGROUND AND OBJECTIVE: In non-coplanar volumetric modulated arc therapy (VMAT) planning, beam trajectory selection (BTS) and fluence map optimization (FMO) are closely coupled. Within a discretized fluence-level appro...
BACKGROUND AND OBJECTIVE: In non-coplanar volumetric modulated arc therapy (VMAT) planning, beam trajectory selection (BTS) and fluence map optimization (FMO) are closely coupled. Within a discretized fluence-level approximation, we formulate this coupling as an integrated mixed-integer nonlinear programming (MINLP) model. METHODS: We solve the resulting MINLP using an alternating optimization strategy. For the BTS subproblem, we construct an ordered layered graph and reformulate dual-trajectory selection as a (2,δ)-partially vertex-disjoint shortest path problem, for which both exact and heuristic algorithms are developed. For the FMO subproblem, we formulate it as a continuous convex optimization problem and solve it using inexact alternating direction method of multipliers (ADMM). RESULTS: We evaluate the proposed framework on four standard matRad benchmark cases (TG119 C-shape, prostate, liver, and head-and-neck) using a simplified discretized non-coplanar angular test platform. Under this experimental setting, the proposed framework suggested favorable organ-at-risk sparing trends relative to the coplanar and greedy baselines while generally maintaining comparable target coverage. CONCLUSIONS: The proposed framework is intended as a fluence-level co-optimization approach for discretized non-coplanar VMAT planning, rather than as a complete clinical VMAT delivery optimization system. The results indicate that explicit coupling of BTS and FMO is feasible under the adopted discretized angular setting. These findings provide a methodological basis for future incorporation of direct aperture optimization and dynamic multileaf collimator trajectory modeling.
An integrated calibration-uncertainty framework for improving the reliability of deep learning models for seizure detection in EEG signals.
Seoni S, Molinari F, Benevieri M … +1 more , Salvi M
Comput Methods Programs Biomed · 2026 Jun · PMID 42308713 · Publisher ↗
BACKGROUND AND OBJECTIVE: Deep learning models have demonstrated strong performance in automated seizure detection from EEG signals. However, these models may produce confident predictions even when incorrect, limiting t...
BACKGROUND AND OBJECTIVE: Deep learning models have demonstrated strong performance in automated seizure detection from EEG signals. However, these models may produce confident predictions even when incorrect, limiting their reliability barrier for clinical adoption. This study proposes an integrated calibration-uncertainty framework to enhance model reliability in EEG-based seizure classification. METHODS: A CNN-BiLSTM model was trained to classify EEG epochs containing epileptic seizure activity. The framework leverages Expected Calibration Error (ECE) to assess global confidence reliability and Monte Carlo Dropout (MCD)-based uncertainty quantification to identify unreliable predictions. For each dropout rate, we evaluated both model calibration and the entropy-based separability between correctly (CC) and misclassified (MC) samples, computed as the Overlap Area between their uncertainty distributions. A multi-objective selection strategy was then used to automatically identify the configuration that best balances these complementary aspects. Finally, a selective classification approach was implemented, using an uncertainty threshold to identify unreliable predictions and defer them for further clinical evaluation. RESULTS: Varying the dropout rate significantly affected both calibration and uncertainty behaviour. The optimal balance was achieved at p = 0.1, yielding the lowest combined ECE and Overlap Area. The selective classification improved accuracy from 91.7% (baseline) to 99.6% while retaining ∼74% of samples, outperforming models optimized for either calibration or uncertainty alone. CONCLUSIONS: The proposed dual perspective framework improves model robustness by integrating global confidence calibration with local uncertainty estimation, representing a practical step toward reliable AI deployment in clinical neurophysiology.
Tonometer- and ultrasound-based pulse wave analysis - a comparative study.
Orter S, Bianchini E, Wassertheurer S … +14 more , Badhwar S, Hametner B, Sinceri S, Bachler M, Ghiadoni L, Boutouyrie P, Aringhieri G, Cosottini M, Guarino D, Faita F, Eugenijus K, Gemignani V, Bruno RM, Mayer CC
Comput Methods Programs Biomed · 2026 Jun · PMID 42302594 · Publisher ↗
BACKGROUND AND OBJECTIVE: Arterial pulse wave analysis (PWA), a tool capable of reflecting regional hemodynamics and wave propagation characteristics, has been primarily based on tonometric measurements, while ultrasound...
BACKGROUND AND OBJECTIVE: Arterial pulse wave analysis (PWA), a tool capable of reflecting regional hemodynamics and wave propagation characteristics, has been primarily based on tonometric measurements, while ultrasound has been used for evaluating local arterial wall mechanics. This study investigates the accuracy of aortic PWA, derived from ultrasound measurements of carotid diameter, in comparison to standard tonometry. METHODS: A sub-dataset of the FUCHSIA study (i.e., patients with fibromuscular dysplasia, matched hypertensives and matched healthy controls) with pairs of carotid ultrasound (MyLab, ESAOTE, Genoa, Italy) and tonometric (SphygmoCor CvMS, Atcor Medical) recordings were used for analyses. Ultrasound longitudinal scans were processed as follows: extraction of diameter curves (Carotid Studio, Quipu); discarding low-quality curves; transformation of diameter to local pressure curves by mathematical models; machine-learning-based transformation from carotid to aortic pulse waves; and finally, calibration of aortic pressure curves. ARCSolver algorithms (AIT) were applied to these ultrasound-based aortic pressure waveforms and the tonometry-based aortic waveforms as reference method to derive PWA parameters (e.g., heart rate (HR), central systolic pressure (cSBP) and augmentation index (AIx)) for pairwise comparison. RESULTS: In total, 74 recordings from 49 patients (51 (SD 15) years; 15 men) were used in this comparative study. Tonometry- and ultrasound-derived HR correlated well (r = 0.88) with no significant bias (mean difference 0.58 (SD 4.9) bpm). Comparable results were obtained for cSBP (r = 0.97, mean difference -3.6 (SD 4.2) mmHg) and AIx (r = 0.77, mean difference -4.2 (SD 9.1) %). CONCLUSIONS: Central PWA parameters obtained from carotid ultrasound-derived aortic pressure curves showed agreement with tonometry-based measurements and results are in line with literature.
A robust machine learning approach for breast cancer subtype classification using relative gene expression order representations.
Uma Kandan S, Abul O
Comput Methods Programs Biomed · 2026 Jun · PMID 42302593 · Publisher ↗
BACKGROUND AND OBJECTIVE: Breast cancer subtype classification is critical for clinical decision-making. It informs prognostic assessment and guides personalized treatment planning. Machine learning has been widely appli...
BACKGROUND AND OBJECTIVE: Breast cancer subtype classification is critical for clinical decision-making. It informs prognostic assessment and guides personalized treatment planning. Machine learning has been widely applied to develop subtype classifiers based on gene expression profiles due to its ability to capture complex molecular patterns. However, technical variation in gene expression data presents a major challenge in building classifiers that maintain robust and generalizable performance across different platforms and patient cohorts. METHODS: We propose a robust machine learning approach leveraging relative gene expression order representations for subtype classification. Two representations: rank- and word2vec embedding-based, were evaluated to capture biological variation across subtypes with minimal cross-sample dependence. These representations were derived from the within-sample relative expression order of PAM50 genes, a well-established 50-gene signature for defining breast cancer subtypes. Machine learning models were trained on these representations and systematically evaluated for robustness and clinical relevance across benchmark datasets. RESULTS: Both rank- and word2vec embedding-based machine learning models demonstrated robust performance during cross-validation on the SCAN-B training set and on SCAN-B internal and semi-external test sets, achieving at least 91% precision and recall. On the fully external TCGA-BRCA dataset, the models accurately classified the majority of subtypes, with at least 95% accuracy. Predicted luminal subtypes also showed clinically meaningful stratification in survival analyses, and both representations effectively preserved biological variation, as reflected in subtype-wise sample segregation in PCA visualizations. In addition, word2vec-derived gene embeddings captured biologically meaningful co-occurring gene clusters. CONCLUSION: Ranks and word2vec embeddings derived from relative gene expression order enable machine learning models to robustly classify breast cancer subtypes. They address technical variability by reducing dependence on cohort-wide normalization, thereby supporting reliable subtype classification in diverse clinical settings.
AutoVARP - A framework for automated reproducible inducibility testing in computational models of cardiac electrophysiology.
Seghetti P, Gsell MAF, Prassl AJ … +2 more , Bishop M, Plank G
Comput Methods Programs Biomed · 2026 Jun · PMID 42296791 · Publisher ↗
BACKGROUND AND OBJECTIVE: Simulations of cardiac electrophysiology (CEP) are gaining momentum beyond basic mechanistic studies, as an approach for supporting clinical decision making. The potential for in silico technolo...
BACKGROUND AND OBJECTIVE: Simulations of cardiac electrophysiology (CEP) are gaining momentum beyond basic mechanistic studies, as an approach for supporting clinical decision making. The potential for in silico technologies observed from the research community is immense, with studies demonstrating significantly improved therapeutical outcome with little to no additional burden for patients. Studies replicating virtually induction protocols in post myocardial infarction patients are among the most reproduced and promise to identify non invasively ablation targets for therapeutical intervention. Two main factors hinder the translation of these technologies from pure research to applications: virtually no reproducibility of results, and lack of standardized procedures. Inspired by a previously published virtual induction study by Arevalo et al. (2016), We address the issues of reproducibility and efficiency providing auto-VARP, a framework for automated virtual arrhythmia inducibility studies, built upon openCARP and the carputils framework. METHODS AND RESULTS: Standardization relies on the previously published forCEPSS framework and is ensured by defining the whole induction study with input files that can be easily shared to ensure reproducibility since the whole pipeline relies on open software. Our approach also ensures numerical efficiency by separating the induction study into four stages: (i) pre-pacing with forCEPSS, (ii) S1 pacing for each steady state, (iii) S2 induction with different extrastimuli, (iv) testing of sustenance of induced reentries. We demonstrate the approach in a large virtual subject cohort to investigate numerical artifacts that may arise when improper setups are provided to perform virtual induction, and additionally showcase auto-VARP in a biventricular mesh. CONCLUSIONS: auto-VARP addresses effectively the current gap in automation and reproducibility of results providing a uniform methodology that can be implemented even by non expert users. auto-VARP is highly scalable and adaptable to markedly different anatomies. Although less flexible than in house implementations it provides automated tools to share setups and does not require re-implementation of any process.
Strategies for incorporating alternative splicing variants in thyroid nodule biopsy classification based on expression levels.
Wilk AM, Łakomiec K, Oczko-Wojciechowska M … +4 more , Kowalska M, Krajewska J, Jarząb B, Student S
Comput Methods Programs Biomed · 2026 Jun · PMID 42296790 · Publisher ↗
BACKGROUND: Accurate diagnosis of thyroid nodules remains a challenge, in cases with indeterminate cytology (Bethesda III and IV). Existing molecular tests leave a percentage of these cases unresolved, leading to unneces...
BACKGROUND: Accurate diagnosis of thyroid nodules remains a challenge, in cases with indeterminate cytology (Bethesda III and IV). Existing molecular tests leave a percentage of these cases unresolved, leading to unnecessary surgeries or delayed treatment. Given that over 90% of genes undergo alternative splicing (AS), this study explores integrating AS data with traditional gene expression profiles for classification. METHODS: Gene expression data from 335 patients were used. HTA2.0 microarrays were preprocessed using two tools for splicing variant identification: the Transcriptome Analysis Console (TAC), which relies on probe sets and junctions, and EventPointer, which focuses on splicing events. The influence of feature selection, dataset and variant identification was tested in a bootstrap procedure. Modification was introduced to deduplicate features for each gene. RESULTS & CONCLUSIONS: The classification quality was strongly influenced by the processing methodology. While the EventPointer pipeline proved more effective for gene-level features due to a custom chip definition file, TAC-generated variants yielded the best bootstrap-based performance, with an overall classification accuracy of 0.938. The model subsequently underwent patient- and sample-level external validation using independent public microarray dataset. Furthermore, we conducted feature verification using RNA-seq data to confirm cross-platform consistency; however, this specific analysis serves as a technical reassessment rather than a full independent classifier validation. Among the 11 selected isoforms were those corresponding to genes known to be significant in thyroid cancer, such as FN1 and LIPH. In thyroid cancer, certain transcript isoforms may be preferentially expressed. Therefore, diagnostic classifiers might benefit from incorporating alternative splicing variants.
Graph sample and aggregate-based deep operator network for accelerating one-dimensional human breathing simulations.
Phung KV, Tran TM, Nguyen QH … +1 more , Choi S
Comput Methods Programs Biomed · 2026 Jun · PMID 42287773 · Publisher ↗
BACKGROUND AND OBJECTIVE: Neural network (NN)-based surrogate models represent a promising tool for accelerating human lung airflow simulations. However, the highly hierarchical branching structure of the lung airway pro...
BACKGROUND AND OBJECTIVE: Neural network (NN)-based surrogate models represent a promising tool for accelerating human lung airflow simulations. However, the highly hierarchical branching structure of the lung airway produces flow rates ranging from liters to microliters per second, which, along with the geometric complexity of the airway, complicates the design of appropriate NN models. METHODS: We introduce a novel model combining Deep Operator Network (DeepONet) with Graph Sample and Aggregate (GraphSAGE), called GraphDeepONet, to accelerate simulations of one-dimensional (1D) human airway models. We also develop a volume-based normalization technique to eliminate the effects of multiscale flow rate variations during model training. We design three different types of models and compare their predictions with a conventional 1D solver as a reference. Model 1 employs a dual-model GraphDeepONet to predict the flow rate and static pressure simultaneously. Model 2 predicts the flow rate first and then computes the static pressure accordingly. Model 3 predicts the flow rates in the acinar regions, hierarchically aggregates them to obtain the remaining flow rates, and then calculates the static pressure, similar to Model 2. RESULTS: Model 3 provided the most accurate predictions for healthy subjects, achieving median (Q1-Q3) L relative errors of 7.0% (6.6-14.4%) in flow rate, 5.9% (5.6-14.3%) in static pressure, and 6.8% (6.3-12.8%) in pleural pressure. These NN models significantly lower computation costs (<1 s) compared with conventional solvers (∼12 min). CONCLUSIONS: The proposed surrogate modeling strategy ensures efficient simulation of airflow in 1D human airways, potentially enabling real-time respiratory analysis and personalized treatment.
FSCL-BC: Federated supervised contrastive learning for breast cancer diagnosis with high sensitivity.
Ahmed F, Sánchez D, Haddi Z … +1 more , Domingo-Ferrer J
Comput Methods Programs Biomed · 2026 Jun · PMID 42285014 · Publisher ↗
BACKGROUND AND OBJECTIVE: Accurate diagnosis of breast cancer in dense breasts requires expert radiologists to examine multiple ultrasound images per patient. This diagnosis procedure is tedious, time-consuming, and pron...
BACKGROUND AND OBJECTIVE: Accurate diagnosis of breast cancer in dense breasts requires expert radiologists to examine multiple ultrasound images per patient. This diagnosis procedure is tedious, time-consuming, and prone to misdiagnosis due to human fatigue. AI-aided diagnosis systems can help alleviate this burden. However, vast amounts of data from multiple hospitals, diverse patient demographics, imaging scanners, and protocols are required to develop accurate, robust, and generalizable AI models. Obtaining such a mixture of data is quite challenging due to privacy concerns, data ownership issues, and regulatory constraints. Moreover, due to subtle visual cues, low resolution, a limited number of labeled samples in hospital datasets, and substantial class imbalance inherent in cancer imaging, deep learning models often overfit to the majority (benign) class. As a result, they struggle to generalize well to unseen data and to achieve high sensitivity, thereby increasing the risk of missed cancer cases. To address these problems, this study aims to develop an accurate AI model for breast cancer prediction from ultrasound images using data from multiple hospitals without requiring data sharing. METHODS: We introduce FSCL-BC, a privacy-preserving method for the diagnosis of breast cancer from ultrasound images with improved sensitivity, integrating supervised contrastive learning within federated learning. This allows hospitals to keep their data on their premises while collaboratively training the AI model, only exchanging the model parameters trained on their private data. RESULTS: Experimental evaluation within a realistic federated setting shows that FSCL-BC achieved substantially higher diagnostic performance - especially sensitivity - than standardized centralized and vanilla federated training, while providing intrinsic privacy protection. On average, compared to the best baseline, FSCL-BC improved sensitivity by 11.5%, Youden's J index by 7.2%, and F1 score by 3.7%. The improvements in MCC and balanced accuracy were more modest, at 2.8% and 2.4%, respectively. CONCLUSION: Improved diagnostic accuracy and enhanced patient privacy preservation make FSCL-BC a promising and practical solution for developing AI models for breast cancer diagnosis from ultrasound images, particularly in real-world, resource-constrained settings.
From black-box to clinical reasoning: CARE-MD a novel framework for Self-eXplainable skin lesions.
Kandpal KC, Verma P
Comput Methods Programs Biomed · 2026 Jun · PMID 42275844 · Publisher ↗
BACKGROUND AND OBJECTIVE: Deep learning models have achieved remarkable diagnostic accuracy in medical imaging, yet their lack of interpretability limits clinical trust and deployment. This study presents CARE-MD (Clinic...
BACKGROUND AND OBJECTIVE: Deep learning models have achieved remarkable diagnostic accuracy in medical imaging, yet their lack of interpretability limits clinical trust and deployment. This study presents CARE-MD (Clinical Algorithm for Reasoning-Enhanced Medical Diagnosis), a self-explainable framework that mirrors the clinical reasoning process to provide transparent, concept-driven, and case-referable diagnosis of skin lesions. METHODS: CARE-MD follows a four-phase reasoning structure: Observation, Interpretation, Reference, and Validation. In Phase 1, attention-guided localization isolates the lesion region using a modified U-Net backbone. Phase 2 employs a Concept Bottleneck Model (CBM) model to predict human-interpretable dermatological concepts. Phase 3 introduces prototype-based reasoning to compare latent features with previously learned prototypes for case-level interpretability. Phase 4 validates explanations through consistency analysis between attention maps, concept activations, and clinician-annotated lesion regions. The framework was evaluated on two publicly available datasets-ISIC 2018 and HAM10000-using metrics such as accuracy, precision, sensitivity, specificity, Dice coefficient, and explanation consistency score. RESULTS: CARE-MD achieved 87.2% accuracy, 86.8% precision, 85.6% sensitivity, and 88.1% specificity on the ISIC 2018 dataset. Cross-dataset testing on HAM10000 showed stable performance with minimal decline across all metrics, indicating strong generalization capability. The attention-guided module improved Dice score by 3% compared with baseline Attention U-Net, while the prototype referencing module enhanced explanation consistency by 6%. Qualitative results confirmed that CARE-MD produces coherent alignment between salient lesion regions, predicted dermatological concepts, and prototype-based references. CONCLUSIONS: CARE-MD provides a structured approach toward aligning deep learning predictions with clinical reasoning by embedding interpretability into each diagnostic phase. While the framework shows improved interpretability and competitive performance relative to the evaluated baselines, it introduces additional architectural complexity and relies on concept-level supervision. These findings suggest that CARE-MD represents a promising direction for self-explainable medical diagnosis, while also highlighting inherent trade-offs between interpretability, model complexity, and annotation requirements. A notable strength of CARE-MD is its unified integration of multiple interpretability mechanisms within a clinically motivated reasoning structure.
Development and clinical usability assessment of a web-based AI platform for automated echocardiography analysis.
Nazari M, Emami H, Rabiei R … +3 more , Rabiee HR, Salari A, Sadr H
Comput Methods Programs Biomed · 2026 Jun · PMID 42263543 · Publisher ↗
BACKGROUND: Echocardiography is the cornerstone of cardiovascular diagnosis, yet its manual interpretation is labor-intensive and prone to inter-observer variability. While Deep Learning (DL) offers expert-level potentia...
BACKGROUND: Echocardiography is the cornerstone of cardiovascular diagnosis, yet its manual interpretation is labor-intensive and prone to inter-observer variability. While Deep Learning (DL) offers expert-level potential, existing models struggle with clinical generalization due to domain shifts and are often limited to single-view analysis, failing to provide the comprehensive assessment required in real-world practice. OBJECTIVES: To overcome these limitations, this study aimed to design, develop, and clinically evaluate EchoAI, a secure, browser-based Clinical Decision Support System to bridge the gap between high-performance DL algorithms and routine echocardiography analysis. METHODS: We developed a secure web-based framework that integrates our previously validated, multi-task UDA-VAE engine capable of simultaneous quantification of Left Ventricular Ejection Fraction (LVEF) and Wall Thickness across multiple standard acoustic windows (A4C, A2C, PLAX). Uniquely, the platform employs a User-Centered Design with a "Human-in-the-loop" workflow, transforming the AI from a "black box" into a transparent assistant that allows physicians to visualize and verify segmentation masks in real-time. RESULTS: A multicenter clinical validation involving 18 cardiologists and residents across an academic hospital and a private cardiac center demonstrated real-time performance with an average processing time of 1.15 s per cycle across diverse ultrasound vendors. The system achieved a strong correlation with expert measurements (r = 0.98, P < 0.001) and a negligible bias of 0.12%. Usability assessment yielded a high overall satisfaction score (6.20/7). Notably, physicians accepted 86% of the AI-generated outputs without modification (84% in the academic setting and 88% in the private sector), confirming the system's robust reliability and cross-domain adaptability. CONCLUSION: EchoAI demonstrates that integrating vendor-agnostic, domain-adaptive AI into an intuitive, interactive web interface effectively bridges the gap between algorithmic capability and clinical adoption. This multicenter approach significantly reduces manual workload while fostering the high level of clinical trust necessary for routine deployment across diverse healthcare settings.
Asian perspectives on multimodal AI in breast cancer: A call for population-specific validation.
Li L, Yao Q, Chen D
Comput Methods Programs Biomed · 2026 Jun · PMID 42259191 · Publisher ↗
Abstract loading — click title to view on PubMed.
Perceptual quality assessment in digital pathology: Modeling diagnostic usability from expert opinions.
Nguyen T, Mahmoudpour S, Courtoy GE … +6 more , Waelput W, Forsyth R, De Vylder J, Diricx B, Vandemeulebroucke J, Schelkens P
Comput Methods Programs Biomed · 2026 Jun · PMID 42250321 · Publisher ↗
Reliable quality assessment in digital pathology is essential to ensure the diagnostic usability of whole slide images (WSIs), as artifacts introduced during tissue preparation and scanning can degrade image quality and...
Reliable quality assessment in digital pathology is essential to ensure the diagnostic usability of whole slide images (WSIs), as artifacts introduced during tissue preparation and scanning can degrade image quality and affect clinical interpretation. In this paper, we propose a framework that combines subjective usability evaluation with an objective no-reference quality assessment method. A dataset was constructed from WSIs of four tissue types (breast, fertility, gastrointestinal, and lung), where pristine patches were systematically degraded using simulated artifacts including blur, contrast, and color variations. A subjective study with eight pathologists was conducted using a five-point diagnostic usability scale, from which Mean Usability Scores (MUS) were derived and statistically validated. An objective metric was then developed based on contrastive learning-driven pseudo-reference generation, followed by a siamese feature extraction and regression model to predict usability. The proposed method shows strong correlation with expert scores and outperforms several existing quality assessment metrics, while demonstrating consistent performance across multiple distortion types and tissue categories. Our proposed model outperforms competing objective metrics, achieving strong consistency with subjective scores with SRCC of 0.945, PLCC of 0.952, and AUC of 0.98 on the benchmark dataset. The proposed objective metric, together with the designed subjective assessment method and the publicly available dataset, provides a reliable framework for expert-aligned quality assessment in digital pathology.
Application of multimodal large language models for HER2/CEN17 FISH signal detection.
Fryszkowski KS, Les T
Comput Methods Programs Biomed · 2026 May · PMID 42248130 · Publisher ↗
BACKGROUND AND OBJECTIVE: Fluorescence in situ hybridisation (FISH) is a reference technique for HER2 gene amplification assessment, yet manual signal counting is labour-intensive and subject to inter-observer variabilit...
BACKGROUND AND OBJECTIVE: Fluorescence in situ hybridisation (FISH) is a reference technique for HER2 gene amplification assessment, yet manual signal counting is labour-intensive and subject to inter-observer variability. This study evaluates the zero and few-shot capability of multimodal large language models for automatic counting of HER2 and CEN17 signals in single-nucleus images. METHODS: A data set of 240 nuclei, categorised by difficulty (simple, complex, and cluster), was extracted from clinical slides acquired at 100× magnification. Three models-Gemini 2.5 Pro, GPT-5, and Claude Sonnet 4.5 were tested using natural language prompts. In the main evaluation setting on original colour images, Gemini 2.5 Pro outperformed the others, achieving a mean absolute error (MAE) of 0.47 signals for simple nuclei and 0.92 for complex cases. RESULTS: The model demonstrated high consistency across repeated runs (median absolute deviation of 0.22) and relied heavily on colour information, as the aggregated MAE worsened to 1.60 on greyscale inputs. Bland-Altman analysis revealed a systematic overcounting bias of 0.39 signals (p<0.001), driven primarily by HER2 (bias 0.75, p<0.001), whereas CEN17 counts showed no significant bias (bias 0.09, p=0.213). CONCLUSIONS: These results indicate that the model is more likely to count background features as additional signals than to miss true signals. Although current error rates suggest they are not yet sufficiently reliable for autonomous clinical decision-making, the results demonstrate that general-purpose multimodal models can achieve sub-signal accuracy without specific training, indicating that this direction is worth pursuing. This work is intended as an exploratory feasibility study examining the reasoning capabilities of MLLMs as zero-shot counting engines.
PMB-NN: Physiology-centered hybrid AI for personalized hemodynamic monitoring from photoplethysmography.
Zhang Y, Fresiello L, Veltink PH … +2 more , Donker DW, Wang Y
Comput Methods Programs Biomed · 2026 Jun · PMID 42248129 · Publisher ↗
BACKGROUND AND OBJECTIVE: Continuous monitoring of blood pressure (BP) and hemodynamic parameters such as peripheral resistance (R) and arterial compliance (C) are critical for early vascular dysfunction detection and th...
BACKGROUND AND OBJECTIVE: Continuous monitoring of blood pressure (BP) and hemodynamic parameters such as peripheral resistance (R) and arterial compliance (C) are critical for early vascular dysfunction detection and therapy optimization. While photoplethysmography (PPG) wearables has gained widespread popularity, existing data-driven methods for BP estimation lack physiological interpretability. METHODS: We advanced our previously proposed physiology-centered hybrid AI method-the Physiological Model-Based Neural Network (PMB-NN)-in arterial hemodynamic monitoring, that unifies deep learning with a 2-element Windkessel based physiological model (PM) parameterized by R and C acting as physics constraints. The PMB-NN model was trained in a subject-specific manner using PPG-derived timing features as inputs, while demographic information was used to infer cardiac output (Q) which is integrated as a prerequisite to solve PM for R and C during training. The model outputs personalized systolic and diastolic BP. The PMB-NN was initially validated on a primary cohort of 10 healthy young adults performing multi-day static and cycling activities to assess day-to-day robustness, benchmarked against established deep learning (DL) models (FCNN, CNN-LSTM, and Transformer) as well as the standalone PM. This core evaluation followed a tripartite framework: (i) estimation accuracy for BP; (ii) physiologically constrained interpretability, reflected by the model's ability to infer R and C; and (iii) physiological plausibility, assessed via the correlation between estimated BP and input timing features. To explore the mechanistic and operational boundaries of the framework, we quantified the impacts of motion artifacts, Q estimation errors, and calibration efficiency, while assessing model transferability from exercise-driven steady states to distinct autonomic challenges across an age-diverse cohort (n=37, 18-88 years old). RESULTS: PMB-NN achieved systolic BP accuracy (median MAE: 7.2 mmHg) comparable to DL benchmarks, despite yielding diastolic performance (median MAE: 3.9 mmHg) lower than DL models. Crucially, however, PMB-NN exhibited substantially higher physiological plausibility than both DL baselines and PM, suggesting that the hybrid architecture unifies and enhances the respective merits of physiological principles and data-driven techniques. Beyond BP, PMB-NN also identified R (median MAE: 0.13 mmHg s/ml) and C (median MAE: 0.19 ml/mmHg) during training, achieving parameter estimation fidelity similar to the standalone PM, demonstrating that the embedded physiological constraints confer interpretability to the hybrid AI framework. Furthermore, extended evaluations confirmed the framework's resilience to moderate motion artifacts (SNR = 15 dB) and identified 5-min calibration as a pragmatic threshold for personalization. While numerical errors increased under autonomic challenges, it captured age-related hemodynamic shifts while maintaining consistent physiological plausibility across the adult lifespan. CONCLUSIONS: These results position PMB-NN as a balanced, physiologically grounded alternative to purely data-driven approaches for daily hemodynamic monitoring.
Real time prediction of instantaneous wave-free ratio based on lumped parameter and neural network models.
Huang S, Liu Y, Liu J … +4 more , Sun H, Li G, Liu J, Li B
Comput Methods Programs Biomed · 2026 May · PMID 42248128 · Publisher ↗
BACKGROUND AND OBJECTIVE: The Instantaneous wave-Free Ratio (iFR) is considered a potential alternative to Fractional Flow Reserve (FFR) in the diagnosis of coronary artery disease, as it does not require vasodilator adm...
BACKGROUND AND OBJECTIVE: The Instantaneous wave-Free Ratio (iFR) is considered a potential alternative to Fractional Flow Reserve (FFR) in the diagnosis of coronary artery disease, as it does not require vasodilator administration. Nonetheless, its clinical use is similarly restricted by the need for invasive measurement. Therefore, there is an urgent need to develop a non-invasive method for calculating iFR, in order to enable rapid and non-invasive diagnosis of myocardial ischemia. METHODS: We propose a physical framework, termed the Lumped Parameter-Neural Network (LPNN), that combines a lumped parameter model (LPM) of coronary arteries with a fully connected neural network (NN)-based stenotic resistance model for accurate prediction of coronary artery resting flow and pressure, and apply it to noninvasive calculation of iFR (LPNN-iFR). In addition, the principle of coronary compensation is also considered in the LPNN model to accurately obtain the resistance boundary conditions at resting, and the personalized aortic pressure and cardiac output are optimized by adding simulated annealing method, thereby saving calculation time. The accuracy of LPNN-iFR was verified in 61 patients measured clinically. RESULTS: LPNN-iFR has a good correlation (r = 0.94) and consistency (bias:0.0, LOA: -0.05 to 0.05) with the measured iFR, and the calculation time for a cycle of specific patients is 6 s. The sensitivity, specificity, positive predictive value, negative predictive value and accuracy of LPNN-iFR were: 85.71%, 93.94%, 92.31%, 88.57%, 90.16%. The area under curve of LPNN-iFR was 0.933. CONCLUSIONS: LPNN-iFR demonstrates good diagnostic performance and short calculation time, proving clinical application value as an alternative non-invasively method for auxiliary preoperative diagnosis of myocardial ischemia.
AI and radiomics for improving the medical workflow for Multiple Sclerosis.
Cipriani A, Polsinelli M, Placidi G
Comput Methods Programs Biomed · 2026 Jun · PMID 42248127 · Publisher ↗
BACKGROUND AND OBJECTIVE: Multiple Sclerosis (MS) is a chronic degenerative disorder that significantly affects the quality of life of patients. Magnetic Resonance Imaging (MRI) has become essential for diagnosis, monito...
BACKGROUND AND OBJECTIVE: Multiple Sclerosis (MS) is a chronic degenerative disorder that significantly affects the quality of life of patients. Magnetic Resonance Imaging (MRI) has become essential for diagnosis, monitoring, and treatment planning. However, variability in data acquisition and interpretation limits reproducibility. This study aims to design a standardized pipeline that integrates Artificial Intelligence (AI) and radiomics to improve precision, reproducibility, and clinical utility in the evaluation of MS. METHODS: We propose a pipeline that incorporates preprocessing to harmonize imaging data, synthesis of missing modalities, automatic segmentation of lesions, and registration with anatomical and connectomic atlases. New radiomic features were extracted to quantify the characteristics of the lesions, their relationship to the nerve tracts, and the affected cortical regions. RESULTS: The proposed pipeline stabilized resolution and contrast variability among different examinations, eliminated rater dependency in lesion segmentation, and enabled the extraction of new consistent radiomics. Indeed, the proposed radiomics captured the position and orientation of the lesions, as well as the involvement of nerves and cortical regions, offering additional information beyond the conventional lesion volume. Experiments demonstrated that the proposed radiomics are robust to inter-rater variability, being six times less sensitive than lesion volume, thus gaining reproducibility, specificity to lesion position/orientation, and precision. CONCLUSIONS: This work introduces a structured and reproducible approach to integrate AI and radiomics into the clinical workflow for MS. By stabilizing imaging data and enabling advanced radiomic analyses, the pipeline supports predictive modeling. These findings suggest promising opportunities for future applications. However, the pipeline was validated on a limited public dataset; therefore, external clinical/radiological validation is required before considering its use in diagnosis or therapy-planning settings.
Robust and Interpretable AI for Acute Appendicitis: A Simulation-to-Clinical Validation Pipeline.
Uddin KN, Njie E, Jin R … +3 more , Ghose P, Zahid A, Jamil HM
Comput Methods Programs Biomed · 2026 May · PMID 42242041 · Publisher ↗
BACKGROUND AND OBJECTIVE: Accurate and timely diagnosis of Acute Appendicitis remains a major challenge in emergency medicine, largely because of its overlapping symptoms and heterogeneous clinical presentations. This st...
BACKGROUND AND OBJECTIVE: Accurate and timely diagnosis of Acute Appendicitis remains a major challenge in emergency medicine, largely because of its overlapping symptoms and heterogeneous clinical presentations. This study introduces the Robust and Interpretable Simulation-Analysis (RISA) framework, a dual-phase methodology designed to evaluate diagnostic classifiers for robustness, generalizability, and interpretability. METHODS: Six supervised learning algorithms (Fisher Discriminant Analysis, Linear Discriminant Analysis, Quadratic Discriminant Analysis, Logistic Regression, Support Vector Machine, and Random Forest) were systematically assessed through 24 factorial simulation scenarios that varied in class imbalance, dimensionality, covariance structure, and nonlinearity of decision boundaries. The simulation phase was followed by clinical validation on two independent datasets: the adult Appendicitis dataset (n=106) and the Regensburg Pediatric Appendicitis (RPA) dataset (n=782). Model interpretability was investigated using Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). RESULTS: In simulation experiments, Linear Discriminant Analysis and Support Vector Machine exhibited balanced sensitivity and specificity, whereas Random Forest displayed the highest resilience under nonlinear and heterogeneous situations. In clinical validation, Linear Discriminant Analysis attained the greatest AUC of 0.908 on the appendicitis dataset and 0.729 on the RPA dataset. For the Appendicitis dataset, Fisher Discriminant Analysis demonstrated optimal sensitivity (1.00), whereas Quadratic Discriminant Analysis attained flawless specificity (1.00). In the RPA cohort, the Support Vector Machine demonstrated the highest overall discrimination (AUC = 0.762) with consistent fold-wise performance. LIME and SHAP consistently identified clinically established biomarkers, including white blood cell count, neutrophil percentage, and body temperature, supporting the medical validity of the model findings. CONCLUSIONS: The RISA framework offers a reproducible and transparent methodology for creating reliable diagnostic artificial intelligence by integrating simulation-based benchmarking with interpretable clinical validation. It connects algorithmic robustness with clinical reliability, facilitating the implementation of explainable machine learning models for the diagnosis of Acute Appendicitis.
About
- Frequency
- Sun
- Papers found
- 200
- RSS feed
- Subscribe