High-field 7T Image Synthesis Using Generative Adversarial Network for Enhancing Epilepsy Diagnosis
Abstract number :
3.361
Submission category :
5. Neuro Imaging / 5A. Structural Imaging
Year :
2024
Submission ID :
437
Source :
www.aesnet.org
Presentation date :
12/9/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Tamjid Imtiaz, MS – University of Pennsylvania
Alfredo Lucas, PhD – University of Pennsylvania
Nishant Sinha, PhD – University of Pennsylvania
Joel Stein, MD, PhD – University of Pennsylvania
Sandhitsu Das, PhD – University of Pennsylvania
Kathryn Davis, MD – University of Pennsylvania
Rationale: Magnetic Resonance Imaging is an efficient tool for precisely localizing the epileptogenic zone in epilepsy patients. However, previous studies showed that approximately 30% of lesional epilepsy cohorts that are mistakenly considered as MRI-negative in the conventional widely available 3-tesla (T) MRI. Epilepsy surgical outcomes are better if a lesion can be identified on MRI. Visualizing these subtle lesions requires a high-field MRI scanner that requires significant capital investment. On the other hand, deep generative models have been widely used for image translation in medical imaging which can learn the mapping from one modality to another. Leveraging this paradigm, we present and assess a novel deep-learning framework that can generate synthetic 7T brain images from the corresponding low-field 3T inputs that can offer potential avenues for enhancing diagnostic precision in epilepsy management.
Methods:
We collected data from a cohort of 30 patients using both standard 3T and 7T scanners at Penn, acquiring T1-weighted images. Subsequently, we devised a framework incorporating a generative adversarial network (GAN) tailored for translating images between the 3T and 7T modalities. We evaluated the synthetic images in regard to the image quality, morphometry and the ability to visualize lesions.
Results:
Synthetic 7T images exhibited visually superior quality compared to their 3T counterparts, with significantly higher normalized cross-correlation (NCC) to actual high-field images for T1-weighted scans (p< 0.001). Additionally, comparative contrast-to-noise ratios between grey and white matter were observed in the synthetic images. Volumetric analysis indicated that critical regions implicated in epilepsy, particularly smaller areas like the hippocampus and thalamus, exhibited minimal deviation from actual 7T measurements. Notably, visual examination of focal cortical dysplasia subjects in synthetic 7T images revealed a significant improvement in contrast compared to the 3T images.
Conclusions: This method creates synthetic high-field images generated from 3T MRI that closely resemble actual 7T images, providing a valuable diagnostic tool in settings lacking access to 7T protocol.
Funding: R01NS116504
Neuro Imaging