AI Model Restores Missing Brain MRI Sequences With High Anatomical Accuracy

A new machine learning model that generates anatomically faithful brain MRI sequences could help overcome a common challenge in clinical imaging—missing or unusable scans—while improving downstream diagnostic accuracy.

Presented Wednesday by researchers from Stony Brook University’s Imaging Informatics for Precision Medicine (IMAGINE) Lab, the study introduces a diffusion-based generative AI framework capable of synthesizing missing brain MRI sequences without hallucinating structures or distorting pathology.

“Baseline methods hallucinate anatomy structure and tumor occurrence,” said Moinak Bhattacharya, PhD candidate and lead presenter. “Our proposed method generates clinically accurate MR sequences.”

In clinical neuroimaging, patients often have incomplete MRI datasets due to movement during scanning, intolerance to long protocols, or contraindications to contrast agents. The absence of sequences like T1, T2, or FLAIR can compromise not only radiologist interpretations but also AI-powered tools, which depend on the full suite of sequences for optimal performance.

To fill the gap, the team built a two-stage generative model. First, they used structural priors—specifically gray and white matter maps—to guide the model with an anatomical reference. Then, they applied a topology-preserving loss function to maintain the shape and consistency of anatomical features in the synthetic images.

“Topology-preserving losses maintain structural consistency, ensuring that generated anatomies remain clinically faithful,” Bhattacharya explained.

In an evaluation of 78 patient cases, the new model outperformed existing generative models in preserving anatomical detail and tumor structure. Radiologists rated the AI-generated images higher for overall image quality, anatomical accuracy, and realistic pathology.

The synthetic images were also able to reflect clinically relevant tumor features, including stratification based on MGMT methylation status—a biomarker that influences response to chemotherapy in glioblastoma.

“This approach has direct clinical implications by enabling robust diagnostic workflows even when one or more MRI sequences are missing,” Bhattacharya said. “It can be integrated into clinical imaging pipelines to enhance diagnostic completeness, support radiologist decision-making and facilitate consistent AI model performance in incomplete imaging scenarios.”

The team is now working on improving the model’s computational efficiency and building trust for clinical deployment. Bhattacharya and his collaborators, including Dr. Gagandeep Singh, a neuroradiologist at Columbia University Irving Medical Center, are actively testing the model’s performance in clinical environments.

The researchers also see broader applications beyond MRI. “This model can help improve the data inadequacies that currently plague a lot of the medical research out there,” Bhattacharya noted, citing potential use in digital pathology where data scarcity is a major obstacle.

Co-authors include Prateek Prasanna, PhD, associate professor of biomedical informatics at Stony Brook, and Annie Singh from the Atal Bihari Vajpayee Institute of Medical Sciences in India.

By combining anatomical awareness with modern generative techniques, the model opens new pathways for restoring diagnostic completeness when imaging data is incomplete—without compromising fidelity or clinical confidence.