Abhirup Banerjee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Abhirup Banerjee has a total of 253 co-authors. Here are the top co-authors by number of shared publications:
- Vicente Grau
University of Oxford
24 shared publications - Marcel Beetz
Department of Engineering Science, Institute of Biomedical Engineering, University of Oxford, Oxford, U.K.
14 shared publications - Pradipta Maji
Indian Statistical Institute
10 shared publications Robin P. Choudhury
Division of Cardiovascular Medicine Radcliffe Department of Medicine University of Oxford UK
10 shared publicationsVicente Grau
Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford
10 shared publicationsMarcel Beetz
Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford
9 shared publications
Latest Publications
- Open Access
- 9 October 2026
- Published by Springer Nature in npj Drug Discovery
- Vol. 3 (1) , 42
TechBio 3.0 closes the drug discovery loop with multimodal AI and generative chemistry
Artificial intelligence (AI) has transformed molecular prediction and design, yet these advances remain only partially integrated with experimental discovery. We define TechBio 3.0 as an emerging paradigm for closed-loop drug discovery that connects multimodal molecular representations, generative chemistry, and automated experimentation. We examine its early preclinical and clinical evidence, identify remaining technical and translational barriers, and argue that this convergence is beginning to define biotechnology’s next era.
- 9 October 2026
- Published by Springer Nature in npj Drug Discovery
- Vol. 3 (1) , 42
Open AccessTechBio 3.0 closes the drug discovery loop with multimodal AI and generative chemistry
Artificial intelligence (AI) has transformed molecular prediction and design, yet these advances remain only partially integrated with experimental discovery. We define TechBio 3.0 as an emerging paradigm for closed-loop drug discovery that connects multimodal molecular representations, generative chemistry, and automated experimentation. We examine its early preclinical and clinical evidence, identify remaining technical and translational barriers, and argue that this convergence is beginning to define biotechnology’s next era.
- Open Access
- 8 October 2026
- Published by MDPI AG in Bioengineering
- Vol. 13 (10) , 1168
BMA-Net: A Bilateral Attention Network with Retinal Domain Transfer-Learning for CIMT-Based Cardiovascular Risk Classification from Fundus Images
Carotid intima-media thickness (CIMT) is an established marker of subclinical atherosclerosis, but its reliance on ultrasound equipment and trained operators limits its suitability for large-scale screening. Retinal fundus photography may offer a more scalable approach to vascular assessment, as it contains retinal vascular features associated with systemic cardiovascular health and can be acquired quickly and non-invasively. This study proposes a two-stage transfer learning framework for CIMT-based cardiovascular risk classification. An attention-enhanced EfficientNet-B4 is first pretrained on the Asia Pacific Tele-Ophthalmology Society 2019 Blindness Detection dataset and subsequently transferred to a weight-sharing Siamese architecture to classify the China Fundus Carotid Intima-Media Thickness dataset. Across ten independent initializations, the proposed model achieved an average macro-F1 score of 77.81% and an area under the receiver operating characteristic curve of 84.08% on the validation set and 78.29% and 85.22%, respectively, on the test set. It achieved overall higher point estimates for performance metrics than those previously reported for the Siamese ResNeXt baseline model with squeeze-and-excitation attention, and demonstrated a reduced disparity in class-wise performance. Grad-CAM analysis further revealed differences in the spatial activation patterns between models trained under different experimental configurations. Overall, these findings support the feasibility of fundus imaging as a scalable and non-invasive approach to CIMT-based cardiovascular risk assessment.
- 8 October 2026
- Published by MDPI AG in Bioengineering
- Vol. 13 (10) , 1168
Open AccessBMA-Net: A Bilateral Attention Network with Retinal Domain Transfer-Learning for CIMT-Based Cardiovascular Risk Classification from Fundus Images
Carotid intima-media thickness (CIMT) is an established marker of subclinical atherosclerosis, but its reliance on ultrasound equipment and trained operators limits its suitability for large-scale screening. Retinal fundus photography may offer a more scalable approach to vascular assessment, as it contains retinal vascular features associated with systemic cardiovascular health and can be acquired quickly and non-invasively. This study proposes a two-stage transfer learning framework for CIMT-based cardiovascular risk classification. An attention-enhanced EfficientNet-B4 is first pretrained on the Asia Pacific Tele-Ophthalmology Society 2019 Blindness Detection dataset and subsequently transferred to a weight-sharing Siamese architecture to classify the China Fundus Carotid Intima-Media Thickness dataset. Across ten independent initializations, the proposed model achieved an average macro-F1 score of 77.81% and an area under the receiver operating characteristic curve of 84.08% on the validation set and 78.29% and 85.22%, respectively, on the test set. It achieved overall higher point estimates for performance metrics than those previously reported for the Siamese ResNeXt baseline model with squeeze-and-excitation attention, and demonstrated a reduced disparity in class-wise performance. Grad-CAM analysis further revealed differences in the spatial activation patterns between models trained under different experimental configurations. Overall, these findings support the feasibility of fundus imaging as a scalable and non-invasive approach to CIMT-based cardiovascular risk assessment.
- 25 September 2026
- Published by Springer Nature
- p. 583-593
Multi-stage NeRF for Efficient 3D Coronary Artery Reconstruction from Two Narrow-Angle Angiographic Projections
- 25 September 2026
- Published by Springer Nature
- p. 583-593
Multi-stage NeRF for Efficient 3D Coronary Artery Reconstruction from Two Narrow-Angle Angiographic Projections
- 25 September 2026
- Published by Springer Nature
- p. 560-571
HeartFormer: Semantic-Aware Dual-Structure Transformers for 3D Four-Chamber Cardiac Point Cloud Reconstruction
- 25 September 2026
- Published by Springer Nature
- p. 560-571
HeartFormer: Semantic-Aware Dual-Structure Transformers for 3D Four-Chamber Cardiac Point Cloud Reconstruction
- Open Access
- 21 June 2026
- Published by Wolters Kluwer Health in Circulation
- Vol. 154 (4) , 316-333
Contrastive Machine Learning to Quantify Hypertensive Multiorgan Damage and Identify New Disease Phenotypes: A Multinational Multimodal Study
BACKGROUND: Hypertension induces structural and functional damage in multiple organs. Evidence of subclinical damage increases risk of vascular events and death but can be difficult to identify in the clinic. We developed a novel machine learning approach that quantifies current hypertension-associated multiorgan damage, mapping progression from health to advanced disease, in a pseudotemporal manner and predicts organ-specific disease progression trajectories. METHODS: We analyzed 566 multimodal imaging and nonimaging variables from 27 099 participants in the UK Biobank imaging substudy to develop a semisupervised contrastive trajectory inference (cTI) framework that models multiorgan alterations associated with hypertension exposure, including heart, brain, kidneys, vasculature, lungs, liver, and metabolic information. Model stability was validated through multiple internal validation steps, and external validity was tested on 5507 participants from the Atherosclerosis Risk in Communities study (ARIC). Clinical relevance was evaluated against existing risk scores and through ability to predict survival and incident multiorgan disease for up to 7 years, across both UK Biobank and ARIC. RESULTS: In the UK Biobank (mean age 63.27±7.48 years; 53.4% women) our global organ damage score (HyperScore) achieved an area under the curve of 0.964 (0.941–0.987) for identification of individuals with severe end-organ disease and robust stability in cross-validation with a mean root mean square error of 0.104±0.084. Survival odds differed significantly across HyperScore stages ( P 0.05) and consistent end-organ and outcome characteristics between ARIC and UK Biobank across HyperTrajectories. CONCLUSIONS: Machine learning–derived global organ damage scores are feasible in hypertension and enable identification of distinct hypertension-associated organ-disease phenotypes. New frameworks for hypertension assessment and monitoring using imaging to derive personalized risk assessment and phenotype-specific intervention may be achievable.
- 21 June 2026
- Published by Wolters Kluwer Health in Circulation
- Vol. 154 (4) , 316-333
Open AccessContrastive Machine Learning to Quantify Hypertensive Multiorgan Damage and Identify New Disease Phenotypes: A Multinational Multimodal Study
BACKGROUND: Hypertension induces structural and functional damage in multiple organs. Evidence of subclinical damage increases risk of vascular events and death but can be difficult to identify in the clinic. We developed a novel machine learning approach that quantifies current hypertension-associated multiorgan damage, mapping progression from health to advanced disease, in a pseudotemporal manner and predicts organ-specific disease progression trajectories. METHODS: We analyzed 566 multimodal imaging and nonimaging variables from 27 099 participants in the UK Biobank imaging substudy to develop a semisupervised contrastive trajectory inference (cTI) framework that models multiorgan alterations associated with hypertension exposure, including heart, brain, kidneys, vasculature, lungs, liver, and metabolic information. Model stability was validated through multiple internal validation steps, and external validity was tested on 5507 participants from the Atherosclerosis Risk in Communities study (ARIC). Clinical relevance was evaluated against existing risk scores and through ability to predict survival and incident multiorgan disease for up to 7 years, across both UK Biobank and ARIC. RESULTS: In the UK Biobank (mean age 63.27±7.48 years; 53.4% women) our global organ damage score (HyperScore) achieved an area under the curve of 0.964 (0.941–0.987) for identification of individuals with severe end-organ disease and robust stability in cross-validation with a mean root mean square error of 0.104±0.084. Survival odds differed significantly across HyperScore stages ( P 0.05) and consistent end-organ and outcome characteristics between ARIC and UK Biobank across HyperTrajectories. CONCLUSIONS: Machine learning–derived global organ damage scores are feasible in hypertension and enable identification of distinct hypertension-associated organ-disease phenotypes. New frameworks for hypertension assessment and monitoring using imaging to derive personalized risk assessment and phenotype-specific intervention may be achievable.
- Open Access
- 15 April 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (4)
Applications of artificial intelligence and computational approaches to imaging for hypertension identification, phenotyping, and outcome prediction: a systematic review
Current hypertension guidelines focus on blood pressure control, but incorporating end-organ imaging could improve understanding of disease manifestations. We undertook a systematic review to evaluate current task-level applications of artificial intelligence (AI) and computational approaches to imaging for hypertension identification, phenotyping, and outcome prediction. A systematic search was conducted across multiple databases up to end of December 2025. Retrieved studies were grouped by AI task, and a thematic qualitative analysis per-task was conducted to evaluate organ-specific findings, AI methodologies, and research gaps. For quantitative synthesis, the I2 statistic derived from Cochran’s Q test was used to assess heterogeneity, and forest plots were generated to visualize effect sizes. The review was registered with PROSPERO (CRD42023427430). The search strategy yielded 48 studies. Thematic analysis categorized the studies into five major tasks, with the majority employing supervised learning for classification processes. Nearly half of the studies focused on the heart. However, paucity of studies performed multi-organ assessment, external validation, and phenotyping or predicting future risk. AI and computational approaches in imaging achieved an overall sensitivity of 0.84 [0.69–0.93] in identifying hypertension from normotension, highest with brain imaging. Sensitivity reached 0.92 [0.90–0.94] in discriminating hypertension from hypertrophic cardiomyopathy. Current research focusses primarily on hypertension prediction using single organ information. While results are promising, datasets remain small with limited external validation. There remains a need for discovery-oriented research to uncover disease heterogeneity, multi-organ phenotypes, and support personalized and targeted interventions.
- 15 April 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (4)
Open AccessApplications of artificial intelligence and computational approaches to imaging for hypertension identification, phenotyping, and outcome prediction: a systematic review
Current hypertension guidelines focus on blood pressure control, but incorporating end-organ imaging could improve understanding of disease manifestations. We undertook a systematic review to evaluate current task-level applications of artificial intelligence (AI) and computational approaches to imaging for hypertension identification, phenotyping, and outcome prediction. A systematic search was conducted across multiple databases up to end of December 2025. Retrieved studies were grouped by AI task, and a thematic qualitative analysis per-task was conducted to evaluate organ-specific findings, AI methodologies, and research gaps. For quantitative synthesis, the I2 statistic derived from Cochran’s Q test was used to assess heterogeneity, and forest plots were generated to visualize effect sizes. The review was registered with PROSPERO (CRD42023427430). The search strategy yielded 48 studies. Thematic analysis categorized the studies into five major tasks, with the majority employing supervised learning for classification processes. Nearly half of the studies focused on the heart. However, paucity of studies performed multi-organ assessment, external validation, and phenotyping or predicting future risk. AI and computational approaches in imaging achieved an overall sensitivity of 0.84 [0.69–0.93] in identifying hypertension from normotension, highest with brain imaging. Sensitivity reached 0.92 [0.90–0.94] in discriminating hypertension from hypertrophic cardiomyopathy. Current research focusses primarily on hypertension prediction using single organ information. While results are promising, datasets remain small with limited external validation. There remains a need for discovery-oriented research to uncover disease heterogeneity, multi-organ phenotypes, and support personalized and targeted interventions.
- Open Access
- 6 April 2026
- Published by Springer Nature in Research Square
Coronary Arteries Segmentation in Invasive X-ray Angiography: A Comprehensive Review and Benchmarking
Coronary artery disease remains the leading cause of morbidity and mortality worldwide, and X-ray coronary angiography (XCA) is the gold standard for its diagnosis and management during real-time cardiac interventions. Accurate segmentation of coronary arteries in XCA is a critical step in quantitative analysis, supporting stenosis detection, treatment planning, and 3D reconstruction. However, segmentation is highly challenging due to overlapping vessels, bone shadows, low contrast, and complex vascular geometry. In this review, we provide the first comprehensive synthesis that systematically categorises and critically evaluates segmentation methods for XCA, covering both classical image processing techniques and emerging machine learning and deep learning approaches. We summarise their evolution, strengths, and limitations, and present benchmarking of advanced deep learning models on two public datasets using Dice score, sensitivity, and precision. The observed performance variability highlights the need for robust algorithms capable of addressing label scarcity, cross-domain generalisability, and interpretability. By consolidating methodological advances and benchmarking evidence, this review offers a foundation to guide future developments in reliable coronary artery segmentation for clinical translation.
- 6 April 2026
- Published by Springer Nature in Research Square
Open AccessCoronary Arteries Segmentation in Invasive X-ray Angiography: A Comprehensive Review and Benchmarking
Coronary artery disease remains the leading cause of morbidity and mortality worldwide, and X-ray coronary angiography (XCA) is the gold standard for its diagnosis and management during real-time cardiac interventions. Accurate segmentation of coronary arteries in XCA is a critical step in quantitative analysis, supporting stenosis detection, treatment planning, and 3D reconstruction. However, segmentation is highly challenging due to overlapping vessels, bone shadows, low contrast, and complex vascular geometry. In this review, we provide the first comprehensive synthesis that systematically categorises and critically evaluates segmentation methods for XCA, covering both classical image processing techniques and emerging machine learning and deep learning approaches. We summarise their evolution, strengths, and limitations, and present benchmarking of advanced deep learning models on two public datasets using Dice score, sensitivity, and precision. The observed performance variability highlights the need for robust algorithms capable of addressing label scarcity, cross-domain generalisability, and interpretability. By consolidating methodological advances and benchmarking evidence, this review offers a foundation to guide future developments in reliable coronary artery segmentation for clinical translation.
- Open Access
- 23 March 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 14
Anatomical basis of sex differences in the electrocardiogram identified by three-dimensional torso-heart imaging reconstruction pipeline
The electrocardiogram (ECG) is used for diagnosis and risk stratification in myocardial infarction (MI). Women have a higher incidence of missed MI diagnosis and complications following infarction, and to address this we aim to provide quantitative information on sex-differences in ECG and torso-ventricular anatomical features and their interdependence. A novel computational automated pipeline is presented enabling the three-dimensional reconstruction of torso-ventricular anatomies for 425 post-MI subjects and 1051 healthy controls from UK Biobank clinical images. Regression models were created relating torso-ventricular and ECG parameters. We found that female hearts were positioned more posteriorly and superiorly than male, and in MI hearts were oriented more horizontally, especially for women. Post-MI women exhibited less QRS prolongation, requiring 27% more prolongation than men to exceed 120ms. Only half of the sex difference in QRS duration was associated with smaller female cavities. Lower STj amplitude in women was striking, associated with smaller ventricles, but also more superior and posterior cardiac position. Post-MI, T wave amplitude and R axis deviations were more strongly associated with posterior and horizontal cardiac positioning in women than in men. Our study highlights the need to quantify sex differences in anatomical features, their implications in ECG interpretation, and the application of clinical ECG thresholds in post-MI.
- 23 March 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 14
Open AccessAnatomical basis of sex differences in the electrocardiogram identified by three-dimensional torso-heart imaging reconstruction pipeline
The electrocardiogram (ECG) is used for diagnosis and risk stratification in myocardial infarction (MI). Women have a higher incidence of missed MI diagnosis and complications following infarction, and to address this we aim to provide quantitative information on sex-differences in ECG and torso-ventricular anatomical features and their interdependence. A novel computational automated pipeline is presented enabling the three-dimensional reconstruction of torso-ventricular anatomies for 425 post-MI subjects and 1051 healthy controls from UK Biobank clinical images. Regression models were created relating torso-ventricular and ECG parameters. We found that female hearts were positioned more posteriorly and superiorly than male, and in MI hearts were oriented more horizontally, especially for women. Post-MI women exhibited less QRS prolongation, requiring 27% more prolongation than men to exceed 120ms. Only half of the sex difference in QRS duration was associated with smaller female cavities. Lower STj amplitude in women was striking, associated with smaller ventricles, but also more superior and posterior cardiac position. Post-MI, T wave amplitude and R axis deviations were more strongly associated with posterior and horizontal cardiac positioning in women than in men. Our study highlights the need to quantify sex differences in anatomical features, their implications in ECG interpretation, and the application of clinical ECG thresholds in post-MI.
- Open Access
An interpretable and explainable neural network to classify sports-related cardiac arrhythmias in professional football athletes
Sudden cardiac death risk is 2-3-fold higher in athletes than in non-athletes. We classify sports-related cardiac arrhythmias using a novel explainability framework comprising data analysis, model interpretability, post-hoc visualisation, and systematic assessment. Two neural networks—one with interpretable sinc convolution and one with standard convolution—were trained on general-population ECGs (PhysioNet, n=88,253, 30 arrhythmias, three continents) and tested on professional footballers (PF12RED, n=161) via domain adaptation for normal sinus rhythm (NSR), sinus bradycardia (SB), incomplete right bundle branch block (IRBBB), and T-wave inversion (TWI). Sinc convolution achieved superior NSR detection (AUROC 0.75 vs 0.70), whilst standard convolution excelled at SB (0.74 vs 0.73), IRBBB (0.66 vs 0.58), and TWI (0.59 vs 0.54). Gradient-weighted Class Activation Mapping revealed that sinc models focus on physiologically relevant ECG segments (the PR interval for NSR/SB and the T wave for TWI). We hypothesise that sinc convolution better captures periodic rhythms but struggles with complex morphological patterns, suggesting architectural choice should align with underlying cardiac pathophysiology. Graphical Abbreviations: AI, artificial intelligence; AUPRC, area under the precision-recall curve; AUROC, area under the receiver operating characteristic curve; Conv, convolution; ECG, electrocardiogram; Grad-CAM, gradient-weighted class activation mapping; IAVB, first-degree atrioventricular block; IRBBB, incomplete right bundle branch block; LAD, left axis deviation; LBBB, left bundle branch block; LVH, left ventricular hypertrophy; NSR, normal sinus rhythm; QT, QT interval; RAD, right axis deviation; RBBB, right bundle branch block; RVH, right ventricular hypertrophy; SA, sinus arrhythmia; SB, sinus bradycardia; TWI, T-wave inversion; xAI, explainable artificial intelligence.
Open AccessAn interpretable and explainable neural network to classify sports-related cardiac arrhythmias in professional football athletes
Sudden cardiac death risk is 2-3-fold higher in athletes than in non-athletes. We classify sports-related cardiac arrhythmias using a novel explainability framework comprising data analysis, model interpretability, post-hoc visualisation, and systematic assessment. Two neural networks—one with interpretable sinc convolution and one with standard convolution—were trained on general-population ECGs (PhysioNet, n=88,253, 30 arrhythmias, three continents) and tested on professional footballers (PF12RED, n=161) via domain adaptation for normal sinus rhythm (NSR), sinus bradycardia (SB), incomplete right bundle branch block (IRBBB), and T-wave inversion (TWI). Sinc convolution achieved superior NSR detection (AUROC 0.75 vs 0.70), whilst standard convolution excelled at SB (0.74 vs 0.73), IRBBB (0.66 vs 0.58), and TWI (0.59 vs 0.54). Gradient-weighted Class Activation Mapping revealed that sinc models focus on physiologically relevant ECG segments (the PR interval for NSR/SB and the T wave for TWI). We hypothesise that sinc convolution better captures periodic rhythms but struggles with complex morphological patterns, suggesting architectural choice should align with underlying cardiac pathophysiology. Graphical Abbreviations: AI, artificial intelligence; AUPRC, area under the precision-recall curve; AUROC, area under the receiver operating characteristic curve; Conv, convolution; ECG, electrocardiogram; Grad-CAM, gradient-weighted class activation mapping; IAVB, first-degree atrioventricular block; IRBBB, incomplete right bundle branch block; LAD, left axis deviation; LBBB, left bundle branch block; LVH, left ventricular hypertrophy; NSR, normal sinus rhythm; QT, QT interval; RAD, right axis deviation; RBBB, right bundle branch block; RVH, right ventricular hypertrophy; SA, sinus arrhythmia; SB, sinus bradycardia; TWI, T-wave inversion; xAI, explainable artificial intelligence.
- Open Access
- 25 February 2026
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 23 (235)
Statistical shape modeling in cardiovascular disease: a narrative review
Cardiovascular diseases (CVDs) remain a leading cause of mortality worldwide. We explore the application of statistical shape modeling (SSM) as a powerful tool in cardiac anatomy assessment, facilitating innovative approaches to diagnosis and treatment. SSM uses advanced mathematical and statistical techniques to understand the geometric properties of anatomical structures across populations. By identifying significant shape parameters, it captures and quantifies subtle variations that may elude traditional approaches. We discuss its evolution, from landmark-based methods to point distribution models for establishing the point-to-point correspondence crucial for accurate shape analysis. We delve into the statistical techniques used to measure shape variability, with a focus on principal component analysis for dimensionality reduction. Key evaluation metrics in the assessment of model performance, such as compactness, generalization and specificity, are reviewed. The clinical utility of SSM across the spectrum of CVDs is examined, covering diagnosis, risk stratification, treatment optimization, follow-up and research applications. Future directions, including the development of multi-label models, integration of deep learning approaches, and spatio-temporal SSM to capture dynamic changes in cardiac geometry, are considered. Through this narrative review, we aim to underscore SSM’s promise as a powerful tool in combating CVDs and advancing personalized medicine, ultimately improving patient outcomes.
- 25 February 2026
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 23 (235)
Open AccessStatistical shape modeling in cardiovascular disease: a narrative review
Cardiovascular diseases (CVDs) remain a leading cause of mortality worldwide. We explore the application of statistical shape modeling (SSM) as a powerful tool in cardiac anatomy assessment, facilitating innovative approaches to diagnosis and treatment. SSM uses advanced mathematical and statistical techniques to understand the geometric properties of anatomical structures across populations. By identifying significant shape parameters, it captures and quantifies subtle variations that may elude traditional approaches. We discuss its evolution, from landmark-based methods to point distribution models for establishing the point-to-point correspondence crucial for accurate shape analysis. We delve into the statistical techniques used to measure shape variability, with a focus on principal component analysis for dimensionality reduction. Key evaluation metrics in the assessment of model performance, such as compactness, generalization and specificity, are reviewed. The clinical utility of SSM across the spectrum of CVDs is examined, covering diagnosis, risk stratification, treatment optimization, follow-up and research applications. Future directions, including the development of multi-label models, integration of deep learning approaches, and spatio-temporal SSM to capture dynamic changes in cardiac geometry, are considered. Through this narrative review, we aim to underscore SSM’s promise as a powerful tool in combating CVDs and advancing personalized medicine, ultimately improving patient outcomes.