AI Fills in Missing MRI Sequences in Brain Tumors

A sophisticated machine learning-based model that trained on brain anatomy improves the accuracy and realism of synthesized brain MRI sequences, according to research presented on Wednesday.

The absence of one or more MRI sequences is a frequent challenge in clinical neuroimaging. Sequences are often missing or unobtainable due to motion artifacts, patient intolerance and contraindications to contrast agents. 

Exacerbating the problem, existing AI-based diagnostic and segmentation tools require all standard sequences for optimal performance.

Researchers at the Imaging Informatics for Precision Medicine (IMAGINE) Lab at Stony Brook University in New York, developed a diffusion model-driven generative AI framework for synthesizing missing brain MRI sequences. Unlike conventional diffusion models that generate images from random noise and lack anatomical fidelity, the new approach can generate anatomically faithful and diagnostically meaningful images.

“Baseline methods hallucinate anatomy structure and tumor occurrence,” said presenter and study coauthor Moinak Bhattacharya, a PhD candidate at Stony Brook. “Our proposed method generates clinically accurate MR sequences.”

Images Have Better Anatomy, Pathology, Quality

The researchers built the model through two stages. They started with anatomical conditioning using structural priors like gray and white matter, then they deployed topology-preserving loss function to help the model achieve state-of-the-art performance. 

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

In an image quality assessment on 78 patients, the model surpassed existing baselines, preserving both brain anatomical details and tumor topology. Quantitative comparative analysis showed that the proposed method achieved better performance on image quality for both the datasets.

“From the radiologists’ evaluation, our generated images have better anatomy, pathology and image quality compared to the baselines,” Bhattacharya said.

The model was able to generate images stratified by MGMT methylation status, an important prognostic indicator because tumors with methylated MGMT are more responsive to chemotherapy. 

“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 researchers plan to further develop the model by leveraging better graphics processing units and more computing power while proving that it is trustworthy enough for clinical implementation. Toward that end, they are working closely with study coauthor and clinical collaborator on the project, Gagandeep Singh, MD, a neuroradiologist and assistant professor of radiology at Colombia University Irving Medical Center in New York City, to prove the model’s value in the clinic.

The model also has applications in domains like digital pathology where there is often not enough data to use to train the models. “This model can help improve the data inadequacies that currently plague a lot of the medical research out there,” Bhattacharya said.

Study coauthors are Prateek Prasanna, PhD, associate professor in the Department of Biomedical Informatics at Stony Brook, and Annie Singh from the Atal Bihari Vajpayee Institute of Medical Sciences in India.

Access the session, “Filling the Gaps: Anatomy-Guided Diffusion Models for Synthesizing Missing MRI Sequences in Brain Tumors,” (W2-PSIN-3) on demand at RSNA.org/MeetingCentral.