Abhijit Das

echo $EMAIL

abhijit.das@mbzuai.ac.ae

cat about.txt

I am a PhD student in Machine Learning at MBZUAI in Abu Dhabi and co-founder & CTO of MedOS. My research connects scientific foundation models, autonomous discovery, and trustworthy AI for personalized medicine.

I work on multimodal learning and world models for science, with an interest in agents that propose, test, and refine scientific hypotheses. I study alignment, interpretability, and uncertainty-aware learning, as well as efficient inference and model distillation.

My recent work explores multimodal medical image analysis, out-of-distribution detection, anatomy-structured representations, and optimization and generalization. I also build open-source tools, including BibCheck for Overleaf.

tail -f news.log

[Jun 2026]New preprint: EnTrust models inter-modal conflict for trustworthy medical imaging.
[Jun 2026]PROTON and Graph-of-Differences — MICCAI 2026.
[May 2026]New preprint on Transformer loss landscapes and weight decay.
[2025]Responsible foundation models in medical imaging — Frontiers in Medicine.
[2024]AnoMed — MICCAI; ProFONet and SEANet — ICPR.

ls research/

I am interested in scientific foundation models, trustworthy medical AI, autonomous discovery, and efficient intelligence. My work spans multimodal learning, uncertainty estimation, medical image analysis, and the foundations of optimization. Recent papers are highlighted.

Modality encoders feed shared, unique and conflict features into diffusion segmentation and uncertainty estimation.

EnTrust: Modeling Inter-Modal Conflict for Trustworthy Multimodal Medical Image Analysis

Dwarikanath Mahapatra, Abhijit Das, Behzad Bozorgtabar, Zongyuan Ge, Sudipta Roy, Deepak Nayak, Mauricio Reyes, Imran Razzak

Preprint 2026

paper / figure / full size

@misc{entrust2026,
  title = {{EnTrust: Modeling Inter-Modal Conflict for Trustworthy Multimodal Medical Image Analysis}},
  author = {Dwarikanath Mahapatra and Abhijit Das and Behzad Bozorgtabar and Zongyuan Ge and Sudipta Roy and Deepak Nayak and Mauricio Reyes and Imran Razzak},
  year = {2026},
  howpublished = {Preprint},
  url = {https://arxiv.org/abs/2606.21384}
}

EnTrust separates shared anatomy, modality-specific cues, and inter-modal conflict. A diffusion-based segmentation model uses these signals to express uncertainty where imaging modalities disagree.

bibtex

@misc{entrust2026,
  title = {{EnTrust: Modeling Inter-Modal Conflict for Trustworthy Multimodal Medical Image Analysis}},
  author = {Dwarikanath Mahapatra and Abhijit Das and Behzad Bozorgtabar and Zongyuan Ge and Sudipta Roy and Deepak Nayak and Mauricio Reyes and Imran Razzak},
  year = {2026},
  howpublished = {Preprint},
  url = {https://arxiv.org/abs/2606.21384}
}

Frozen vision-language encoders connect to an online prototype bank, a confidence gate and adaptive out-of-distribution scoring.

PROTON: Prototype-Based Test-Time Online OOD Detection for Medical VLMs

Abhijit Das, Nichula Wasalathilaka, Yifan Lu, Adinath Dukre, Dwarikanath Mahapatra, Shadab Khan, Imran Razzak

MICCAI 2026

paper / figure / full size

@misc{proton2026,
  title = {{PROTON: Prototype-Based Test-Time Online OOD Detection for Medical VLMs}},
  author = {Abhijit Das and Nichula Wasalathilaka and Yifan Lu and Adinath Dukre and Dwarikanath Mahapatra and Shadab Khan and Imran Razzak},
  year = {2026},
  howpublished = {MICCAI},
  url = {https://arxiv.org/abs/2606.20913}
}

PROTON maintains an online prototype bank from high-confidence test predictions and combines prototype distance with concept-matching scores to detect distribution shifts.

bibtex

@misc{proton2026,
  title = {{PROTON: Prototype-Based Test-Time Online OOD Detection for Medical VLMs}},
  author = {Abhijit Das and Nichula Wasalathilaka and Yifan Lu and Adinath Dukre and Dwarikanath Mahapatra and Shadab Khan and Imran Razzak},
  year = {2026},
  howpublished = {MICCAI},
  url = {https://arxiv.org/abs/2606.20913}
}

Anatomical graph construction, graph encoding and soft correspondence connect image representations to difference alignment.

Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-Identification

Nichula Wasalathilaka, Abhijit Das, Imran Razzak, Dwarikanath Mahapatra

MICCAI 2026

paper / figure / full size

@misc{graph2026,
  title = {{Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-Identification}},
  author = {Nichula Wasalathilaka and Abhijit Das and Imran Razzak and Dwarikanath Mahapatra},
  year = {2026},
  howpublished = {MICCAI},
  url = {https://arxiv.org/abs/2606.21368}
}

Graph-of-Differences represents images as graphs of named anatomical regions, matches corresponding nodes, and aligns anatomy-level differences with the global image representation.

bibtex

@misc{graph2026,
  title = {{Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-Identification}},
  author = {Nichula Wasalathilaka and Abhijit Das and Imran Razzak and Dwarikanath Mahapatra},
  year = {2026},
  howpublished = {MICCAI},
  url = {https://arxiv.org/abs/2606.21368}
}

A Transformer block with residual connections and annotated attention, normalization and feed-forward parameter groups.

Weight-Decay Turns Transformer Loss Landscapes Villani: Functional-Analytic Foundations for Optimization and Generalization

Abhijit Das, Sayantan Dutta

Preprint 2026

paper / figure / full size

@misc{landscape2026,
  title = {{Weight-Decay Turns Transformer Loss Landscapes Villani: Functional-Analytic Foundations for Optimization and Generalization}},
  author = {Abhijit Das and Sayantan Dutta},
  year = {2026},
  howpublished = {Preprint},
  url = {https://arxiv.org/abs/2605.06599}
}

A functional-analytic study of Transformer objectives with weight decay, investigating coercive loss geometry, convergence of noisy optimization, and generalization bounds.

bibtex

@misc{landscape2026,
  title = {{Weight-Decay Turns Transformer Loss Landscapes Villani: Functional-Analytic Foundations for Optimization and Generalization}},
  author = {Abhijit Das and Sayantan Dutta},
  year = {2026},
  howpublished = {Preprint},
  url = {https://arxiv.org/abs/2605.06599}
}

Data sources, model training, ethical considerations, downstream medical tasks and deployment in a foundation-model workflow.

Ethical Framework for Responsible Foundational Models in Medical Imaging

Debesh Jha, Gorkem Durak, Abhijit Das, Jasmer Sanjotra, Onkar Susladkar, Suramyaa Sarkar, Ashish Rauniyar, et al.

Frontiers in Medicine 2025

paper / figure / full size

@misc{ethics2025,
  title = {{Ethical Framework for Responsible Foundational Models in Medical Imaging}},
  author = {Debesh Jha and Gorkem Durak and Abhijit Das and Jasmer Sanjotra and Onkar Susladkar and Suramyaa Sarkar and Ashish Rauniyar and others},
  year = {2025},
  howpublished = {Frontiers in Medicine},
  url = {https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2025.1544501/full}
}

An ethical framework for medical imaging foundation models, discussing transparency, fairness, privacy, accountability, and responsible deployment.

bibtex

@misc{ethics2025,
  title = {{Ethical Framework for Responsible Foundational Models in Medical Imaging}},
  author = {Debesh Jha and Gorkem Durak and Abhijit Das and Jasmer Sanjotra and Onkar Susladkar and Suramyaa Sarkar and Ashish Rauniyar and others},
  year = {2025},
  howpublished = {Frontiers in Medicine},
  url = {https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2025.1544501/full}
}

Teacher-student learning, the scale-invariant bottleneck and confidence-guided pseudo-label optimization in AnoMed.

Confidence-guided Semi-supervised Learning for Generalized Lesion Localization in X-ray Images

Abhijit Das, Vandan Gorade, Komal Kumar, Snehashis Chakraborty, Dwarikanath Mahapatra, Sudipta Roy

MICCAI 2024

paper / code / figure / full size

@misc{lesion2024,
  title = {{Confidence-guided Semi-supervised Learning for Generalized Lesion Localization in X-ray Images}},
  author = {Abhijit Das and Vandan Gorade and Komal Kumar and Snehashis Chakraborty and Dwarikanath Mahapatra and Sudipta Roy},
  year = {2024},
  howpublished = {MICCAI},
  url = {https://papers.miccai.org/miccai-2024/154-Paper3485.html}
}

A semi-supervised approach to lesion localization in X-ray images, using confidence-aware pseudo-labeling to learn from labeled and unlabeled examples.

bibtex

@misc{lesion2024,
  title = {{Confidence-guided Semi-supervised Learning for Generalized Lesion Localization in X-ray Images}},
  author = {Abhijit Das and Vandan Gorade and Komal Kumar and Snehashis Chakraborty and Dwarikanath Mahapatra and Sudipta Roy},
  year = {2024},
  howpublished = {MICCAI},
  url = {https://papers.miccai.org/miccai-2024/154-Paper3485.html}
}

Support and query embeddings pass through distance learning and feature optimization with variance, invariance and covariance regularization.

ProFONet: Prototypical Feature Space Optimized Network for Few-shot Classification

Abhijit Das, Vandan Gorade, Debesh Jha, Koushik Biswas, Pethuru Raj, Ulas Bagci

ICPR 2024

paper / code / figure / full size

@misc{profonet2024,
  title = {{ProFONet: Prototypical Feature Space Optimized Network for Few-shot Classification}},
  author = {Abhijit Das and Vandan Gorade and Debesh Jha and Koushik Biswas and Pethuru Raj and Ulas Bagci},
  year = {2024},
  howpublished = {ICPR},
  url = {https://link.springer.com/chapter/10.1007/978-3-031-78183-4_25}
}

ProFONet optimizes prototypical representations for few-shot classification using variance, invariance, and covariance regularization to improve class separation.

bibtex

@misc{profonet2024,
  title = {{ProFONet: Prototypical Feature Space Optimized Network for Few-shot Classification}},
  author = {Abhijit Das and Vandan Gorade and Debesh Jha and Koushik Biswas and Pethuru Raj and Ulas Bagci},
  year = {2024},
  howpublished = {ICPR},
  url = {https://link.springer.com/chapter/10.1007/978-3-031-78183-4_25}
}

A CT denoising encoder-decoder joins spatial and spectral residual units across skip connections.

SEANet: Rethinking Skip-Connections Design in Encoder-Decoder Networks via Synergistic Spatial-Spectral Fusion for LDCT Denoising

Abhijit Das, Vandan Gorade, Dwarikanath Mahapatra, Sudipta Roy

ICPR 2024

paper / code / figure / full size

@misc{seanet2024,
  title = {{SEANet: Rethinking Skip-Connections Design in Encoder-Decoder Networks via Synergistic Spatial-Spectral Fusion for LDCT Denoising}},
  author = {Abhijit Das and Vandan Gorade and Dwarikanath Mahapatra and Sudipta Roy},
  year = {2024},
  howpublished = {ICPR},
  url = {https://link.springer.com/chapter/10.1007/978-3-031-78198-8_29}
}

SEANet studies skip connections that combine spatial and spectral information in encoder-decoder networks for low-dose CT denoising.

bibtex

@misc{seanet2024,
  title = {{SEANet: Rethinking Skip-Connections Design in Encoder-Decoder Networks via Synergistic Spatial-Spectral Fusion for LDCT Denoising}},
  author = {Abhijit Das and Vandan Gorade and Dwarikanath Mahapatra and Sudipta Roy},
  year = {2024},
  howpublished = {ICPR},
  url = {https://link.springer.com/chapter/10.1007/978-3-031-78198-8_29}
}

PAM-UNet encoder-decoder architecture, progressive attention gate and feature maps before and after attention.

PAM-UNet: Shifting Attention on Region of Interest in Medical Images

Abhijit Das, Debesh Jha, Vandan Gorade, Koushik Biswas, Hongyi Pan, Zheyuan Zhang, Daniela P. Ladner, Yury Velichko, Amir Borhani, Ulas Bagci

IEEE EMBC 2024

paper / figure / full size

@misc{pam2024,
  title = {{PAM-UNet: Shifting Attention on Region of Interest in Medical Images}},
  author = {Abhijit Das and Debesh Jha and Vandan Gorade and Koushik Biswas and Hongyi Pan and Zheyuan Zhang and Daniela P. Ladner and Yury Velichko and Amir Borhani and Ulas Bagci},
  year = {2024},
  howpublished = {IEEE EMBC},
  url = {https://arxiv.org/abs/2405.01503}
}

PAM-UNet uses progressive attention to focus on regions of interest in medical image segmentation.

bibtex

@misc{pam2024,
  title = {{PAM-UNet: Shifting Attention on Region of Interest in Medical Images}},
  author = {Abhijit Das and Debesh Jha and Vandan Gorade and Koushik Biswas and Hongyi Pan and Zheyuan Zhang and Daniela P. Ladner and Yury Velichko and Amir Borhani and Ulas Bagci},
  year = {2024},
  howpublished = {IEEE EMBC},
  url = {https://arxiv.org/abs/2405.01503}
}

Feature aggregation, channel and spatial attention, multi-scale detection and an example detected polyp.

Enhancing Colonoscopy Outcomes with DAPoDet-based AI for Real-time Sessile Serrated Polyp Detection

Abhijit Das, Debesh Jha, Nikhil Tomar, Neethi Dasu, Mark Geissler, Dayang Wang, Mena Bakhit, Kirti Dasu, Tyler Berzin, Ulas Bagci

DDW · Conference abstract 2024

paper / figure / full size

@misc{dapodet2024,
  title = {{Enhancing Colonoscopy Outcomes with DAPoDet-based AI for Real-time Sessile Serrated Polyp Detection}},
  author = {Abhijit Das and Debesh Jha and Nikhil Tomar and Neethi Dasu and Mark Geissler and Dayang Wang and Mena Bakhit and Kirti Dasu and Tyler Berzin and Ulas Bagci},
  year = {2024},
  howpublished = {DDW · Conference abstract},
  url = {https://www.giejournal.org/article/S0016-5107(24)00898-8/fulltext}
}

A DDW conference abstract investigating real-time AI-assisted detection of sessile serrated polyps in colonoscopy video.

bibtex

@misc{dapodet2024,
  title = {{Enhancing Colonoscopy Outcomes with DAPoDet-based AI for Real-time Sessile Serrated Polyp Detection}},
  author = {Abhijit Das and Debesh Jha and Nikhil Tomar and Neethi Dasu and Mark Geissler and Dayang Wang and Mena Bakhit and Kirti Dasu and Tyler Berzin and Ulas Bagci},
  year = {2024},
  howpublished = {DDW · Conference abstract},
  url = {https://www.giejournal.org/article/S0016-5107(24)00898-8/fulltext}
}

ls side_quests/

A different perspective

A rotating four-dimensional cube, projected into three dimensions. An interactive experiment in geometry.

BibCheck for Overleaf — Reference verification with source-backed corrections.
ProFONet — Few-shot classification with optimized prototypical representations.

cat contact.txt

For research collaborations and opportunities: abhijit.das@mbzuai.ac.ae.