Low-cost neurological diagnostics, disease progression prediction, quantitative measurements, and autonomous medical research
ReMIND (https://remind.bu.edu/) is a vision-language framework to make brain MRI analysis more affordable, predict disease progression, measure brain changes and volumes, and accelerate autonomous medical research.
Our goal is to make brain MRI analysis more affordable, measure changes in brain structure over time, predict neurological disease progression, and support autonomous medical research. We will build on ReMIND (https://www.medrxiv.org/content/10.64898/2026.03.30.26349106), a vision-language framework with 35 billion parameters but only 3 billion activated per token, making it more efficient to run. We will extend ReMIND's agentic framework to use tools such as FastSurfer for brain segmentation and volumetric measurements. We also plan to investigate whether combining brain MRI with longitudinal clinical data can improve diagnosis and disease progression prediction. Building on our previous work on PodGPT (https://www.nature.com/articles/s44385-025-00022-0) and Medical Agentic RAG (https://www.sciencedirect.com/science/article/pii/S1386505626000791), we will integrate large-scale scientific literature retrieval to support autonomous medical research. We aim to make our models, code, and web application publicly available.
Our previous results provide a strong foundation for this project. ReMIND achieved 90.7–96.7% balanced accuracy on closed-ended MRI questions across four external clinical sites. Our contrastive knowledge distillation method also reduced the rate of unsupported answers from 99.70% to 1.52% when the required imaging was missing. In our previous Medical Agentic RAG work, we achieved 82.98% accuracy on USMLE Step 1 and 86.24% on Step 2, outperforming GPT-4 on both benchmarks. Building on these results, we will evaluate whether ReMIND can maintain diagnostic accuracy while reducing inference time and cost per MRI examination. We will assess the accuracy of FastSurfer-based brain volume measurements, evaluate longitudinal changes, and measure disease progression prediction using independent clinical datasets. For autonomous medical research, we will evaluate evidence retrieval accuracy and the quality of AI-generated answers against reference evidence. We will compare our results with existing methods and report both performance improvements and limitations.
We request $100,000 for a 12-month project, including $40,000 for research personnel to develop and evaluate ReMIND, $30,000 for GPU computing to support model training and inference, $10,000 for data acquisition and annotation, and $10,000 for developing and evaluating agentic tools for brain volume measurements, disease progression prediction, and autonomous medical research. The remaining $10,000 will cover indirect and administrative costs.
Our team at Boston University brings together researchers with different backgrounds in radiology, computer science, computer engineering, and statistics. Our project lead, Diala Lteif, is a postdoctoral researcher with extensive experience processing and analyzing millions of MRI images, with publications in Nature Medicine and Human Brain Mapping. I focus on AI model training and system development, with publications at EMNLP and IEEE T-NNLS, over 1,000 citations, and a ranking among the top 1.5% of developers worldwide on GitRanks. Subhrangshu Bit, a Ph.D. candidate in Computer Science, brings expertise in statistics, clinical validation, and U.S. patent preparation. Ria Sonalker recently developed an agentic AI pipeline for quantitative brain MRI measurements. Together, we developed ReMIND and evaluated its performance across multiple clinical datasets, including four external clinical sites. We have a highly collaborative research environment, and each team member brings complementary expertise and a strong commitment to the project.
The main risks are limited access to longitudinal clinical data, differences in MRI acquisition protocols across hospitals, and challenges in generalizing our models to unseen patients. Disease progression prediction may also prove more difficult than diagnosis, particularly when patient follow-up data are incomplete. Integrating quantitative imaging tools and autonomous research agents could introduce additional technical challenges. If the project falls short, we may not achieve the accuracy and reliability needed for clinical applications. However, we will report our findings, identify the limitations, and release our models and code to support future research.
$20,000 from MIT-Anthropic AI for Science Study and Anthropic's Claude Science AI for Science cohort
There are no bids on this project.
Comments
No comments yet. Sign in to create one!