Marius George Linguraru
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Marius George Linguraru has a total of 392 co-authors. Here are the top co-authors by number of shared publications:
- Antonio Reyes PorrasPerez
Department of Biostatistics and Informatics, University of Colorado Anschutz Medical Campus, Aurora, 80045, CO, USA; Department of Pediatric Plastic & Reconstructive Surgery, Children's Hospital Colorado, Aurora, 80045, CO, USA; Department of Neurological Surgery, The University of Chicago, Chicago, 60637, IL, USA
10 shared publications Syed Muhammad Anwar
Children’s National Hospital
10 shared publicationsZhifan Jiang
Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Hospital, Washington, DC, USA
9 shared publicationsAbhijeet Parida
Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Hospital, Washington, DC, USA
7 shared publicationsAustin Tapp
Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Hospital, Washington, DC, USA
7 shared publicationsMaría J. Ledesma-Carbayo
Universidad Politécnica de Madrid and CIBER-BBN, ISCIII, Madrid, Spain
6 shared publications
Latest Publications
- Open Access
- 12 November 2026
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 5 (11) , e0001813
Quantitative facial phenotyping and age-associated differences in Sickle cell disease using interpretable machine learning
Sickle cell disease (SCD) is a common inherited genetic disorder and contributor to global childhood mortality and morbidity, particularly in Sub-Saharan Africa. Craniofacial and skeletal changes have been described in SCD, but the subtle soft-tissue facial patterns and their variation have not been quantitatively characterized using machine-learning methods. We analyzed 308 front-facing smartphone photographs from 154 participants with electrophoresis-confirmed SCD and 154 age- and sex-matched controls aged from 5 months to 19 years recruited in Kinshasa, Democratic Republic of the Congo. Predefined geometric and texture-based facial descriptors were extracted from standardized images using manually annotated landmarks. An interpretable linear support vector machine with recursive feature elimination was evaluated using leave-one-out cross-validation, with features ranked according to their relevance for distinguishing participants with SCD from controls. Cross-sectional age- and sex- stratified analyses evaluated variation in selected features The facial analysis algorithm identified eight geometric and six texture discriminative features that were significantly different between the cohorts. A combination of six geometric features contributed most consistently to internal discrimination: distance between medial and lateral canthi, angle at nasal ala, distance from nasion to philtrum, distance from medial canthi to the columella, distance from columella to the lower lip, and distance between nasal alae. The model achieved an AUROC of 0.843 [95% CI: 0.798,0.887], and an accuracy of 0.795 [95% CI: 0.750,0.840]. Between-group difference was limited in the youngest age groups and were more pronounced in several older age groups. These findings demonstrate preliminary evidence that SCD is associated with quantifiable morphometric patterns. External, longitudinal, and clinically characterized studies are needed to determine whether these features may serve as reproducible digital biomarkers within a broader supportive disease-assessment framework.
- 12 November 2026
- Published by Public Library of Science (PLoS) in PLOS Digital Health
- Vol. 5 (11) , e0001813
Open AccessQuantitative facial phenotyping and age-associated differences in Sickle cell disease using interpretable machine learning
Sickle cell disease (SCD) is a common inherited genetic disorder and contributor to global childhood mortality and morbidity, particularly in Sub-Saharan Africa. Craniofacial and skeletal changes have been described in SCD, but the subtle soft-tissue facial patterns and their variation have not been quantitatively characterized using machine-learning methods. We analyzed 308 front-facing smartphone photographs from 154 participants with electrophoresis-confirmed SCD and 154 age- and sex-matched controls aged from 5 months to 19 years recruited in Kinshasa, Democratic Republic of the Congo. Predefined geometric and texture-based facial descriptors were extracted from standardized images using manually annotated landmarks. An interpretable linear support vector machine with recursive feature elimination was evaluated using leave-one-out cross-validation, with features ranked according to their relevance for distinguishing participants with SCD from controls. Cross-sectional age- and sex- stratified analyses evaluated variation in selected features The facial analysis algorithm identified eight geometric and six texture discriminative features that were significantly different between the cohorts. A combination of six geometric features contributed most consistently to internal discrimination: distance between medial and lateral canthi, angle at nasal ala, distance from nasion to philtrum, distance from medial canthi to the columella, distance from columella to the lower lip, and distance between nasal alae. The model achieved an AUROC of 0.843 [95% CI: 0.798,0.887], and an accuracy of 0.795 [95% CI: 0.750,0.840]. Between-group difference was limited in the youngest age groups and were more pronounced in several older age groups. These findings demonstrate preliminary evidence that SCD is associated with quantifiable morphometric patterns. External, longitudinal, and clinically characterized studies are needed to determine whether these features may serve as reproducible digital biomarkers within a broader supportive disease-assessment framework.
- Open Access
- 10 October 2026
- Published by MDPI AG in Preprints.org
Artificial Intelligence for Medical Image Analysis in Low-Resource Settings: A Systematic Review
Artificial intelligence (AI) for medical image analysis is increasingly being developed for low-resource settings, where specialist expertise, imaging equipment, labelled data, and computing resources may be limited. We systematically reviewed 189 studies published between January 2021 and January 2026 to examine how these constraints relate to imaging modalities, clinical applications, and AI methods in low-resource settings. Each study was also assessed against 15 criteria adapted from the FUTURE-AI principles for trustworthy AI. Shortages of clinical expertise were the most common constraint, and portable or lower-cost imaging was widely used. Transfer learning was the most frequent technical strategy (59%), followed by data augmentation (54%) and efficient model design (36%). Medical-image pretraining and self-supervised learning were each used in 4% of studies, and model compression in 3%. Applications were concentrated in retinal disease, pulmonary disease, antenatal care, parasitic diseases, and cervical screening. Beyond these technical and clinical patterns, the trustworthy AI assessment showed a clear gap between documenting how systems were developed and demonstrating how they performed under relevant local conditions. Subgroup performance was evaluated in 24% of studies, intended users or realistic working conditions in 25%, and model explanations with intended users in 4%. Consistent with this gap, among 15 studies that reported the same model and metric during development and in the local setting, 12 performed worse locally. Future research should prioritise representative local evaluation, intended-user assessment, demonstrated resource suitability, and reproducible research practices.
- 10 October 2026
- Published by MDPI AG in Preprints.org
Open AccessArtificial Intelligence for Medical Image Analysis in Low-Resource Settings: A Systematic Review
Artificial intelligence (AI) for medical image analysis is increasingly being developed for low-resource settings, where specialist expertise, imaging equipment, labelled data, and computing resources may be limited. We systematically reviewed 189 studies published between January 2021 and January 2026 to examine how these constraints relate to imaging modalities, clinical applications, and AI methods in low-resource settings. Each study was also assessed against 15 criteria adapted from the FUTURE-AI principles for trustworthy AI. Shortages of clinical expertise were the most common constraint, and portable or lower-cost imaging was widely used. Transfer learning was the most frequent technical strategy (59%), followed by data augmentation (54%) and efficient model design (36%). Medical-image pretraining and self-supervised learning were each used in 4% of studies, and model compression in 3%. Applications were concentrated in retinal disease, pulmonary disease, antenatal care, parasitic diseases, and cervical screening. Beyond these technical and clinical patterns, the trustworthy AI assessment showed a clear gap between documenting how systems were developed and demonstrating how they performed under relevant local conditions. Subgroup performance was evaluated in 24% of studies, intended users or realistic working conditions in 25%, and model explanations with intended users in 4%. Consistent with this gap, among 15 studies that reported the same model and metric during development and in the local setting, 12 performed worse locally. Future research should prioritise representative local evaluation, intended-user assessment, demonstrated resource suitability, and reproducible research practices.
- Open Access
- 1 August 2026
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 127
Diverse Concept Modeling (DiCoMX): Domain-specific self-supervised pre-training strategy for chest X-ray foundation models
- 1 August 2026
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 127
Open AccessDiverse Concept Modeling (DiCoMX): Domain-specific self-supervised pre-training strategy for chest X-ray foundation models
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
In Utero Imaging with Fetal MRI: Differentiating Patterns of Cranial and Brain Development
Structured Abstract Objective: Although the cranium and brain both grow rapidly during fetal development, tools for visualizing fetal skull separately from the brain have been limited. With increasing availability of fetal magnetic resonance imaging (MRI), we aimed to determine its feasibility for measuring cranial development and characterize how cranial and brain develop across gestation. Methods: This secondary analysis used MRI data from 50 pregnant women with singleton, typically developing pregnancies from 24 to 38.42 weeks’ gestation, collected as part of a longitudinal study of fetal and placental development. Biparietal (BPD) and occipitofrontal (OFD) diameters of the cranium and brain were manually delineated using standard software. Head circumference (HC) was calculated as HC = 1.62*(BPD + OFD), and cephalic index (CI) as CI = (BPD/OFD)*100. Total brain volume was determined as published previously.19 Manual segmentation of the fetal cranium was completed across gestation, and linear regression modeled relations between fetal skull measurements and brain volume. Results: Cranial measurements were successfully obtained from fetal MRI, demonstrating the feasibility of reliably delineating skull BPD and OFD across gestation. BPD ranged from 46.96-94.84mm and OFD ranged from 70.19-113.5mm. Total brain volume showed a positive linear relationship with head circumference, indicating MRI-based cranial measures track brain growth. Cephalic index wasn’t associated with brain volume and did not change with gestational age. Conclusions: Our study demonstrates MRI effectively distinguishes cranial and brain development during the fetal period. Head circumference reflects brain volume, whereas cephalic index does not. Differentiating skull and brain development enhances understanding of fetal neurodevelopment and supports improved prenatal monitoring.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessIn Utero Imaging with Fetal MRI: Differentiating Patterns of Cranial and Brain Development
Structured Abstract Objective: Although the cranium and brain both grow rapidly during fetal development, tools for visualizing fetal skull separately from the brain have been limited. With increasing availability of fetal magnetic resonance imaging (MRI), we aimed to determine its feasibility for measuring cranial development and characterize how cranial and brain develop across gestation. Methods: This secondary analysis used MRI data from 50 pregnant women with singleton, typically developing pregnancies from 24 to 38.42 weeks’ gestation, collected as part of a longitudinal study of fetal and placental development. Biparietal (BPD) and occipitofrontal (OFD) diameters of the cranium and brain were manually delineated using standard software. Head circumference (HC) was calculated as HC = 1.62*(BPD + OFD), and cephalic index (CI) as CI = (BPD/OFD)*100. Total brain volume was determined as published previously.19 Manual segmentation of the fetal cranium was completed across gestation, and linear regression modeled relations between fetal skull measurements and brain volume. Results: Cranial measurements were successfully obtained from fetal MRI, demonstrating the feasibility of reliably delineating skull BPD and OFD across gestation. BPD ranged from 46.96-94.84mm and OFD ranged from 70.19-113.5mm. Total brain volume showed a positive linear relationship with head circumference, indicating MRI-based cranial measures track brain growth. Cephalic index wasn’t associated with brain volume and did not change with gestational age. Conclusions: Our study demonstrates MRI effectively distinguishes cranial and brain development during the fetal period. Head circumference reflects brain volume, whereas cephalic index does not. Differentiating skull and brain development enhances understanding of fetal neurodevelopment and supports improved prenatal monitoring.
- 14 July 2026
- Published by SAGE Publications in The Cleft Palate Craniofacial Journal
Machine Learning-Based Identification of Craniosynostosis From Nonsynostotic Cranial Deformities
Objective: Craniosynostosis, the premature fusion of cranial sutures, affects 5.2 cases per 10,000 live births and can result in impaired neurodevelopment if not treated early, often necessitating surgery. The absence of efficient screening tools for pediatricians during routine visits contributes to unnecessary referrals, treatment delays, and increased healthcare costs. Through machine learning (ML)-based identification of single-suture craniosynostosis and the affected suture, we aim to assist pediatric providers in the early differentiation of synostotic and nonsynostotic cranial conditions using head volume analysis. Design/Methods: Head volumes were collected via photogrammetry or CT. Volumes represented as 3D points were augmented to address data imbalance, and statistical shape modeling was used alongside Principal Component Analysis and an XGBoost ML model to predict head shapes. Three models were trained and evaluated using 10-fold cross-validation: Model-1 identifying synostosis from nonsynostotic shapes, Model-2 identifying the synostotic suture, Model-3 classifying across synostotic and nonsynostotic head shapes. Participants: Five hundred seventy-one subjects (158 normal, 281 deformational plagiocephaly/brachycephaly, 132 single-suture craniosynostosis—including metopic, sagittal, coronal, and lambdoid). Results: Model-1 identified craniosynostosis from nonsynostotic deformities and healthy controls with an average accuracy of 97.8 ± 1.8%, sensitivity of 98.1 ± 2.3%, and specificity of 96.9 ± 1.1%. Model-2 performed with an average sensitivity of 86.1 ± 12.3% and a specificity of 96.8 ± 2.7%. Model-3 had average sensitivity of 76.8 ± 15.9% and specificity of 97.8 ± 1.7% Conclusion: We developed and evaluated an ML approach that reliably differentiates between synostotic and nonsynostotic head shapes, and distinguishes between types of synostotic sutures, including metopic, sagittal, coronal, and lambdoid.
- 14 July 2026
- Published by SAGE Publications in The Cleft Palate Craniofacial Journal
Machine Learning-Based Identification of Craniosynostosis From Nonsynostotic Cranial Deformities
Objective: Craniosynostosis, the premature fusion of cranial sutures, affects 5.2 cases per 10,000 live births and can result in impaired neurodevelopment if not treated early, often necessitating surgery. The absence of efficient screening tools for pediatricians during routine visits contributes to unnecessary referrals, treatment delays, and increased healthcare costs. Through machine learning (ML)-based identification of single-suture craniosynostosis and the affected suture, we aim to assist pediatric providers in the early differentiation of synostotic and nonsynostotic cranial conditions using head volume analysis. Design/Methods: Head volumes were collected via photogrammetry or CT. Volumes represented as 3D points were augmented to address data imbalance, and statistical shape modeling was used alongside Principal Component Analysis and an XGBoost ML model to predict head shapes. Three models were trained and evaluated using 10-fold cross-validation: Model-1 identifying synostosis from nonsynostotic shapes, Model-2 identifying the synostotic suture, Model-3 classifying across synostotic and nonsynostotic head shapes. Participants: Five hundred seventy-one subjects (158 normal, 281 deformational plagiocephaly/brachycephaly, 132 single-suture craniosynostosis—including metopic, sagittal, coronal, and lambdoid). Results: Model-1 identified craniosynostosis from nonsynostotic deformities and healthy controls with an average accuracy of 97.8 ± 1.8%, sensitivity of 98.1 ± 2.3%, and specificity of 96.9 ± 1.1%. Model-2 performed with an average sensitivity of 86.1 ± 12.3% and a specificity of 96.8 ± 2.7%. Model-3 had average sensitivity of 76.8 ± 15.9% and specificity of 97.8 ± 1.7% Conclusion: We developed and evaluated an ML approach that reliably differentiates between synostotic and nonsynostotic head shapes, and distinguishes between types of synostotic sutures, including metopic, sagittal, coronal, and lambdoid.
- 1 May 2026
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- Vol. 8 (3) , e250902
Multi-Institutional Annotated Multiparametric MRI Dataset of Pediatric High-Grade Gliomas
The BraTS-PEDs dataset provides the largest, multi-institutional, publicly available collection of annotated multiparametric MRI scans of pediatric high-grade gliomas, enabling standardized benchmarking and development of AI tools in pediatric neuro-oncology.
- 1 May 2026
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- Vol. 8 (3) , e250902
Multi-Institutional Annotated Multiparametric MRI Dataset of Pediatric High-Grade Gliomas
The BraTS-PEDs dataset provides the largest, multi-institutional, publicly available collection of annotated multiparametric MRI scans of pediatric high-grade gliomas, enabling standardized benchmarking and development of AI tools in pediatric neuro-oncology.
- 1 April 2026
- Published by Springer Nature
- Vol. 16376, 237-247
Improving Pre-trained Adult Glioma Segmentation Models Using only Post-processing Techniques
Gliomas are the most common malignant brain tumors in adults and are among the most lethal. Despite aggressive treatment, the median survival rate is less than 15 months. Accurate multiparametric MRI (mpMRI) tumor segmentation is critical for surgical planning, radiotherapy, and disease monitoring. While deep learning models have improved the accuracy of automated segmentation, large-scale pre-trained models generalize poorly and often underperform, producing systematic errors such as false positives, label swaps, and slice discontinuities in slices. These limitations are further compounded by unequal access to GPU resources and the growing environmental cost of large-scale model training. In this work, we propose adaptive post-processing techniques to refine the quality of glioma segmentations produced by large-scale pretrained models developed for various types of tumors. We demonstrated the techniques in multiple BraTS 2025 segmentation challenge tasks, with the ranking metric improving by 14.9 % for the sub-Saharan Africa challenge and 0.9% for the adult glioma challenge. This approach promotes a shift in brain tumor segmentation research from increasingly complex model architectures to efficient, clinically aligned post-processing strategies that are precise, computationally fair, and sustainable.
- 1 April 2026
- Published by Springer Nature
- Vol. 16376, 237-247
Improving Pre-trained Adult Glioma Segmentation Models Using only Post-processing Techniques
Gliomas are the most common malignant brain tumors in adults and are among the most lethal. Despite aggressive treatment, the median survival rate is less than 15 months. Accurate multiparametric MRI (mpMRI) tumor segmentation is critical for surgical planning, radiotherapy, and disease monitoring. While deep learning models have improved the accuracy of automated segmentation, large-scale pre-trained models generalize poorly and often underperform, producing systematic errors such as false positives, label swaps, and slice discontinuities in slices. These limitations are further compounded by unequal access to GPU resources and the growing environmental cost of large-scale model training. In this work, we propose adaptive post-processing techniques to refine the quality of glioma segmentations produced by large-scale pretrained models developed for various types of tumors. We demonstrated the techniques in multiple BraTS 2025 segmentation challenge tasks, with the ranking metric improving by 14.9 % for the sub-Saharan Africa challenge and 0.9% for the adult glioma challenge. This approach promotes a shift in brain tumor segmentation research from increasingly complex model architectures to efficient, clinically aligned post-processing strategies that are precise, computationally fair, and sustainable.
- 1 April 2026
- Published by Springer Nature
- Vol. 16376, 456-467
Adaptable Segmentation Pipeline for Diverse Brain Tumors with Radiomic-Guided Subtyping and Lesion-Wise Model Ensemble
Robust and generalizable segmentation of brain tumors on multi-parametric magnetic resonance imaging (MRI) remains difficult because tumor types differ widely. The BraTS 2025 Lighthouse Challenge benchmarks segmentation methods on diverse high-quality datasets of adult and pediatric tumors: multi-consortium international pediatric brain tumor segmentation (PED), preoperative meningioma tumor segmentation (MEN), meningioma radiotherapy segmentation (MEN-RT), and segmentation of pre- and post-treatment brain metastases (MET). We present a flexible, modular, and adaptable pipeline that improves segmentation performance by selecting and combining state-of-the-art models and applying tumor- and lesion-specific processing before and after training. Radiomic features extracted from MRI help detect tumor subtype, ensuring a more balanced training. Custom lesion-level performance metrics determine the influence of each model in the ensemble and optimize post-processing that further refines the predictions, enabling the workflow to tailor every step to each case. On the BraTS testing sets, our pipeline achieved performance comparable to top-ranked algorithms across multiple challenges. These findings confirm that custom lesion-aware processing and model selection yield robust segmentations yet without locking the method to a specific network architecture. Our method has the potential for quantitative tumor measurement in clinical practice, supporting diagnosis and prognosis.
- 1 April 2026
- Published by Springer Nature
- Vol. 16376, 456-467
Adaptable Segmentation Pipeline for Diverse Brain Tumors with Radiomic-Guided Subtyping and Lesion-Wise Model Ensemble
Robust and generalizable segmentation of brain tumors on multi-parametric magnetic resonance imaging (MRI) remains difficult because tumor types differ widely. The BraTS 2025 Lighthouse Challenge benchmarks segmentation methods on diverse high-quality datasets of adult and pediatric tumors: multi-consortium international pediatric brain tumor segmentation (PED), preoperative meningioma tumor segmentation (MEN), meningioma radiotherapy segmentation (MEN-RT), and segmentation of pre- and post-treatment brain metastases (MET). We present a flexible, modular, and adaptable pipeline that improves segmentation performance by selecting and combining state-of-the-art models and applying tumor- and lesion-specific processing before and after training. Radiomic features extracted from MRI help detect tumor subtype, ensuring a more balanced training. Custom lesion-level performance metrics determine the influence of each model in the ensemble and optimize post-processing that further refines the predictions, enabling the workflow to tailor every step to each case. On the BraTS testing sets, our pipeline achieved performance comparable to top-ranked algorithms across multiple challenges. These findings confirm that custom lesion-aware processing and model selection yield robust segmentations yet without locking the method to a specific network architecture. Our method has the potential for quantitative tumor measurement in clinical practice, supporting diagnosis and prognosis.
- 1 April 2026
- Published by Springer Nature
- p. 68-79
Mechanistic Learning with Guided Diffusion Models to Predict Spatio-Temporal Brain Tumor Growth
- 1 April 2026
- Published by Springer Nature
- p. 68-79
Mechanistic Learning with Guided Diffusion Models to Predict Spatio-Temporal Brain Tumor Growth
- Open Access
- 17 February 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 315
Unlocking 3D baby face photogrammetry: Multi-view BabyMorph reconstruction from uncalibrated photographs
- 17 February 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 315
Open AccessUnlocking 3D baby face photogrammetry: Multi-view BabyMorph reconstruction from uncalibrated photographs