Next Generation Brain MRI Segmentation for Ultrahighfield MRI
Automatic image segmentation in brain magnetic resonance imaging is essential for various applications, as it is a prerequisite for upstream tasks in computational neuroimaging analysis. Most notably, highly precise segmentation of the cortical gray matter is required for depth-specific fMRI studies, where even minor errors can significantly impact subsequent analyses.
While current applications are generally successful on MR images acquired in the clinical setting at 1.5–3.0 T with a resolution of 0.9–1.2 mm, they become unreliable when applied to ultra-high resolution images and, in particular, to data acquired at 9.4 T magnetic field strength.
This has motivated us to develop a segmentation method capable of handling images at ultra-high resolution ≤ 0.6 mm and field strengths 1.5–9.4 T. Introducing FLEXseg, our method uses an adversarial game for flexible domain adaptation of convolutional neural networks in the context of brain MRI segmentation.
For validation, we meticulously curated a database of 9.4T images with manually corrected labels approved by expert neuroradiologists. Our results demonstrate that FLEXseg outperforms existing methods both quantitatively and qualitatively.
This pioneering 3D segmentation not only overcomes challenges posed by ultra-high field strengths but also paves the way for novel brain studies at 9.4 T, providing a level of accuracy previously unattainable.
Steiglechner, Mahler, Wang, Scheffler, Lohmann
Segmenting Ultrahighfield MRI data using neural networks