Ongoing research · Stanford AIDE Lab
Clinical-report-conditioned 3D CT synthesis
Rushil Srirambhatla, Akshay Chaudhari, David Larson · Stanford AIDE Lab
Fine-tuned three MAISI diffusion U-Nets with ClinicalT5 cross-attention, comparing impression-only, full-report, and MedGemma-structured reports on matched study splits. Built preprocessing for 25,489 Merlin CT studies.
Key finding: Structured conditioning improved BiomedCLIP text–image alignment by ~42%; fixed-noise counterfactual experiments revealed limitations in anatomical control.
BMC Medical Imaging · Pending acceptance
Benchmarking Automated CT Brain Extraction Methods Across 5,000 Scans
Rushil Srirambhatla, Nathan Walborn, Daniel Hanley, John Muschelli
Developed Robust-CTBET and benchmarked seven brain-extraction methods on 5,062 CT scans from 499 patients at 78 sites, assessing registration, volumetrics, and downstream utility.
Key finding: Robust-CTBET outperformed six competing methods, achieving Dice ≈ 0.98 and a 0.30% scan-level failure rate.
ECCV 2026 AI4M3D workshop · Accepted
Memory-Efficient Skip-Connection-Free 3D Autoencoder for Whole-Brain MRI Reconstruction
Rushil Srirambhatla, Helmet T. Karim, Jacques-Yves Campion
Designed a skip-connection-free 3D autoencoder with a 512-dimensional bottleneck and pixel-shuffle decoder for full-volume T1-weighted MRI reconstruction, evaluated across 40,841 scans.
Key finding: 19.6% less peak GPU memory, 10.4% faster forward/backward execution, and 13.4% fewer parameters versus the no-pixel-shuffle baseline.
Manuscript submitted · Co-first author
Deep Learning-Derived Brain Age from 40,000+ MRI Scans Reveals Disease-Specific Aging Trajectories
Jacques-Yves Campion* and Rushil Srirambhatla*, Carmen Andreescu, et al. · *Equal contribution
Developed brain-age prediction pipelines using harmonized 3D MRI and studied disease-associated deviations in predicted brain age across multi-cohort clinical imaging data.
Key finding: 3.14-year held-out test MAE in the brain-age prediction study.
2026 medRxiv preprint · Under revision
Altered T1w/T2w-FLAIR Ratio in White Matter Hyperintensities as an Indicator of Structural Integrity Loss
Rushil Srirambhatla, Jacques-Yves Campion, Thomas Desmidt, et al.
Integrated LST-AI lesion segmentation, ANTs preprocessing, and relative T1w/T2w-FLAIR measures in 576 participants to investigate Alzheimer’s disease, vascular dementia, and amyloid biomarkers.
Key finding: White matter hyperintensity (WMH) T1w/T2w-FLAIR ratios were significantly higher in individuals with Alzheimer’s disease and vascular dementia compared with those with subjective cognitive impairment, suggesting disease-associated alterations in white matter tissue integrity.
2026 · medRxiv preprint · MGH / Harvard
Cortical Thickness Signatures of High vs. Low Cognitive Performers in the ADNI Dataset
Rushil Srirambhatla, Yue Hong, Yiwen Zhang, David H. Salat
Analyzed FreeSurfer-derived cortical thickness across 68 regions in relation to cognition using age-stratified regression, ANCOVA, and FDR correction.
Key finding: Executive-function associations were most extensive among adults aged 80 and older.
2024 · Published commentary · The American Journal of Bioethics
The Black Box Dilemma: Challenges in Human-AI Collaboration in ML-CDSS
Rishab Jain, Rushil Srirambhatla, John Kessler, Ram Goel
A commentary on interpretability and human–AI collaboration in machine-learning-based clinical decision support.
2022 · medRxiv preprint
Prediction of COVID-19 Diagnosis from Healthy and Pneumonia CT Scans Using Convolutional Neural Networks
Rushil Srirambhatla, Helmet T. Karim
Developed an interpretable CNN-based approach to classifying COVID-19 from CT imaging, incorporating lung-focused preprocessing and activation-map visualization.
