Imaging biomarkers for oncology
Turning CT, MRI, and PET into quantitative biomarkers with radiomics and foundation-model features, to support oncology drug development.
Aneesh Rangnekar
Senior AI Scientist, Computational Radiology
AstraZeneca · Boston, MA
At AstraZeneca, I am responsible for developing radiology imaging biomarkers in the Radiology & Theranostics team, turning CT, MRI, and PET into quantitative readouts that help select indications, detect early signs of biological activity, and stratify patients for oncology trials. My research focuses on making vision and multimodal foundation models reliable enough for that setting: robustness to distribution shift, uncertainty, and out-of-distribution detection.
Previously, I was a Research Fellow at Memorial Sloan Kettering Cancer Center with Dr. Harini Veeraraghavan. I received my Ph.D. in Imaging Science from Rochester Institute of Technology, advised by Dr. Matthew Hoffman and co-advised by Dr. Christopher Kanan and Dr. Emmett Ientilucci.
News
New preprint: Adaptive computation with conformal certification, uncertainty-aware sparse inference with risk guarantees for large scientific images.
Joined AstraZeneca in Boston as a Senior AI Scientist in Computational Radiology.
SWIFT, parameter-efficient CT-to-MRI transfer for rectal cancer segmentation, accepted at the MICCAI 2026 MIART workshop; code is out.
Tumor-aware augmentation for CT-to-MRI transfer accepted to Biomedical Signal Processing and Control; code is out.
New preprint: TREX predicts rectal cancer regrowth from longitudinal endoscopy, matching attending surgeons' accuracy.
Research themes
3D and multimodal foundation models
Self-supervised pretraining of 3D transformers on large CT collections, extended to MRI and PET with language supervision.
Knowing when a model is wrong
Lightweight out-of-distribution detection and uncertainty estimates that keep models safe across scanners, protocols, and hospitals.
Selected publications
Google Scholar →
Tumor-anchored deep feature random forests for out-of-distribution detection in lung cancer segmentation
Transactions on Machine Learning Research (TMLR), 2026
Summary
RF-Deep is a lightweight, plug-and-play OOD detector that improves the safety of automated tumor segmentation from 3D CT. It trains random forests on hierarchical deep features from the segmentation model's own backbone, with no changes to the model, to flag imaging outliers. Across 1,916 scans it outperformed existing methods, exceeding 93.5 AUROC on two near-OOD datasets and reaching 99.0 AUROC on two far-OOD datasets.
Tumor-aware augmentation with task-guided attention analysis improves rectal cancer segmentation from magnetic resonance images
Biomedical Signal Processing and Control, 2026
Summary
We identified two interacting failure modes when CT-pretrained transformers are transferred to MRI rectal cancer segmentation: wasted tokens from zero-padding and ineffective feature adaptation. Tumor-aware augmentation and anisotropic cropping fixed both, reaching 91.1% detection with SMIT, 88.7% with Swin UNETR, and 90.3% with VoCo on 247 test cases.
Parameter-efficient pretrained-CT-to-MRI transfer for rectal cancer segmentation: Performance–calibration trade-offs
MICCAI Workshop on Medical Image AI in Radiation Therapy (MIART), 2026
Summary
SWIFT adapts a Swin V2 encoder, pretrained on 10,444 public 3D CT volumes, to T2-weighted MRI. Compressing the decoder cut parameters by 70% while raising tumor detection from 89.9% to 93.9%; LoRA kept similar accuracy with 14.6% of the trainable parameters, and a LoRA-decoder ensemble gave the best calibration.
brat: Aligned multi-view embeddings for brain MRI analysis
Winter Conference on Applications of Computer Vision (WACV), 2026
Summary
brat (Brain Report Alignment Transformer) is trained on 75k MRI–report pairs with a pairwise view-alignment and a diversity-promoting loss. It outperforms prior methods on image–text retrieval, tumor segmentation, and Alzheimer's classification, and enables high-quality automatic report generation from MRIs.
Transformer-based cardiac substructure segmentation from contrast and non-contrast computed tomography for radiotherapy planning
Physics and Imaging in Radiation Oncology, 2026
Summary
A pretrained transformer with a convolutional decoder, fine-tuned with a balanced mix of contrast and non-contrast CTs, approached full-data performance with 64% fewer training scans, generalized from lung to breast cancer patients better than nnU-Net, and produced dose metrics equivalent to manual delineations.
Self-supervised learning improves robustness of deep learning lung tumor segmentation models to CT imaging differences
Medical Physics, 2025
Summary
Large-scale self-supervised pretraining on uncurated, in-the-wild CT made lung tumor segmentation more robust: fine-tuned Swin transformers consistently outperformed self-pretrained Swin and ViT counterparts across varied acquisition protocols.
Semantic segmentation with active semi-supervised learning
Winter Conference on Applications of Computer Vision (WACV), 2023
arXivPublicationCodePosterSlides
Summary
S4AL combines semi-supervised and active learning so that only the most informative image regions are annotated. It reached over 95% of full-dataset performance with under 17% of pixel annotations on CamVid and Cityscapes.
AeroRIT: A new scene for hyperspectral image analysis
IEEE Transactions on Geoscience and Remote Sensing (TGRS), 2020
Summary
AeroRIT is an aerial hyperspectral dataset, with buildings and cars as well as vegetation and roads, designed for training CNNs for scene understanding, with benchmarks across several architectures.
All other publications
- Dual cross-attention siamese transformer for rectal tumor regrowth assessment in watch-and-wait endoscopy J Tapias Gomez, D Kanata, A Rangnekar, C Lee, J Garcia-Aguilar, JJ Smith, H Veeraraghavan · ISBI, 2026
- Random forest-based out-of-distribution detection for robust lung cancer segmentation A Rangnekar, H Veeraraghavan · SPIE Medical Imaging, 2026
- External evaluation of a deep learning-based algorithm for automated lung tumor and esophagus segmentation on CBCT Z Zhang, D Hobbis, J Jiang, A Rangnekar, CMS Choi, M Waters, et al. · Biomedical Physics & Engineering Express, 2026
- Topological data analysis visualization for interpretable assessment of AI contouring quality CMS Choi, A Rangnekar, J Jiang, H Veeraraghavan · SPIE Medical Imaging, 2026
- Out-of-domain data generalization using parallel-route coherent manifold-mixup for prostate cancer segmentation J Simeth, A Rangnekar, V Brennan, S McBride, N Tyagi, H Veeraraghavan · SPIE Medical Imaging, 2026
- Co-distilled attention guided masked image modeling with noisy teacher for self-supervised learning on medical images J Jiang, A Rangnekar, H Veeraraghavan · MIDL, 2025
- Quantifying uncertainty in lung cancer segmentation with foundation models applied to mixed-domain datasets A Rangnekar, N Nadkarni, J Jiang, H Veeraraghavan · SPIE Medical Imaging, 2025
- Swin transformers are robust to distribution and concept drift in endoscopy-based longitudinal rectal cancer assessment J Tapias Gomez, A Rangnekar, H Williams, H Thompson, J Garcia-Aguilar, JJ Smith, H Veeraraghavan · SPIE Medical Imaging, 2025
- Assessing endoscopic response in locally advanced rectal cancer treated with total neoadjuvant therapy: Development and validation of a highly accurate convolutional neural network H Williams, H Thompson, C Lee, A Rangnekar, J Tapias Gomez, M Widmar, et al. · Annals of Surgical Oncology, 2024
- Automatic segmentation for magnetic resonance imaging guided individual elective lymph node irradiation in head and neck cancer patients FCJ Reinders, MHF Savenije, M de Ridder, M Maspero, PAH Doornaert, CHJ Terhaard, CPJ Raaijmakers, K Zakeri, NY Lee, E Aliotta, A Rangnekar, H Veeraraghavan, MEP Philippens · Physics and Imaging in Radiation Oncology, 2024
- 3D Swin transformer for partial medical auto segmentation A Rangnekar, J Jiang, H Veeraraghavan · MICCAI FLARE Challenge, 2023
- DDDAS-based remote sensing M Hoffman, A Rangnekar, Z Mulhollan, A Vodacek · Handbook of Dynamic Data Driven Applications Systems, Vol. 2, 2023
- Semantic segmentation with active semi-supervised representation learning A Rangnekar, C Kanan, M Hoffman · BMVC, 2022
- Semi-supervised hyperspectral object detection challenge results – PBVS 2022 A Rangnekar, Z Mulhollan, A Vodacek, M Hoffman, A Sappa, E Blasch, et al. · CVPR Workshops, 2022
- SpecAL: Towards active learning for semantic segmentation of hyperspectral imagery A Rangnekar, E Ientilucci, C Kanan, M Hoffman · DDDAS, 2022
- Fine-tuning for one-look regression vehicle counting in low-shot aerial datasets A Rangnekar, Y Yao, M Hoffman, A Divakaran · ICPR, 2021
- Uncertainty estimation for semantic segmentation of hyperspectral imagery A Rangnekar, E Ientilucci, C Kanan, M Hoffman · DDDAS, 2020
- Occlusion detection for dynamic adaptation Z Mulhollan, A Rangnekar, A Vodacek, M Hoffman · DDDAS, 2020
- Calibrated vehicle paint signatures for simulating hyperspectral imagery Z Mulhollan, A Rangnekar, T Bauch, M Hoffman, A Vodacek · CVPR Workshops, 2020
- Tracking in aerial hyperspectral videos using deep kernelized correlation filters B Uzkent, A Rangnekar, M Hoffman · IEEE TGRS, 2018
- Aerial vehicle tracking by adaptive fusion of hyperspectral likelihood maps B Uzkent, A Rangnekar, M Hoffman · CVPR Workshops, 2017
Experience
-
AstraZeneca
2026 – presentSenior AI Scientist, Computational Radiology · Boston, MA
Imaging-derived biomarkers for oncology drug development (Radiology & Theranostics).
-
Memorial Sloan Kettering Cancer Center
2022 – 2026Research Fellow, AI for Medical Imaging · New York, NY
Self-supervised foundation models and OOD-detection pipelines for radiology across multiple hospitals; LoRA/DoRA fine-tuning of ~100M-parameter imaging encoders that cut training compute by ~30% and trainable parameters by ~90%.
-
Rochester Institute of Technology
2016 – 2022Graduate Research Assistant · Rochester, NY
Self- and semi-supervised learning for hyperspectral imagery; released the AeroRIT and RooftopHSI datasets.
-
SRI International
Jun – Nov 2018Research Intern · Princeton, NJ
Continual self-supervised learning for aerial object counting.
-
Conduent Labs
Jun – Aug 2017Research Intern · Rochester, NY
GAN-based data augmentation for vehicle occupancy estimation.
Education
- Ph.D., Imaging Science
Rochester Institute of Technology
2015 – 2022 - M.S., Electrical Engineering
Rochester Institute of Technology
2013 – 2015 · degree conferred 2017
Service
Reviewer for CVPR, ICCV, ECCV, MICCAI, Medical Physics, IEEE TIP, and IEEE TGRS.
Toolkit
PyTorch, JAX, Hugging Face, MONAI, Docker, Slurm, Weights & Biases.