Fast automatic segmentation of thalamic nuclei from MP2RAGE acquisition at 7 Tesla

TL;DR

This study demonstrates that the THOMAS algorithm can automatically segment thalamic nuclei from MP2RAGE MRI sequences at 7 Tesla, providing comparable accuracy to specialized WMn-MPRAGE images, thus broadening its clinical and research utility.

Contribution

The paper adapts and validates the THOMAS segmentation method for routine MP2RAGE images, enabling accurate thalamic nuclei segmentation without specialized sequences.

Findings

High segmentation accuracy for large nuclei (dice > 0.85)

Improved performance using synthesized WMn contrast

Equivalent segmentation quality to WMn-MPRAGE images

Abstract

Purpose: Thalamic nuclei are largely invisible in conventional MRI due to poor contrast. Thalamus Optimized Multi-Atlas Segmentation (THOMAS) provides automatic segmentation of 12 thalamic nuclei using white-matter-nulled (WMn) MPRAGE sequence at 7T. Application of THOMAS to Magnetization Prepared 2 Rapid Gradient Echo (MP2RAGE) sequence acquired at 7T has been investigated in this study. Methods: 8 healthy volunteers and 5 pediatric-onset multiple sclerosis patients were recruited at the Children's Hospital of Philadelphia and scanned at Siemens 7T with WMn-MPRAGE and multi-echo MP2RAGE (ME-MP2RAGE) sequences. White-matter-nulled contrast was synthesized (MP2-SYN) from T1 maps from ME-MP2RAGE sequence. Thalamic nuclei were segmented using THOMAS joint label fusion algorithm from WMn-MPRAGE and MP2-SYN datasets. THOMAS pipeline was modified to use majority voting to segment the bias corrected MP2-UNI images. Thalamic nuclei from MP2-SYN and MP2-UNI images were evaluated against corresponding nuclei obtained from WMn-MPRAGE images using dice coefficients, volume similarity indices (VSI) and distance between centroids. Results: For MP2-SYN, dice > 0.85 and VSI > 0.95 was achieved for the 5 larger nuclei and dice > 0.6 and VSI > 0.7 was achieved for the 7 smaller nuclei. The dice and VSI were slightly higher whilst the distance between centroids were smaller for MP2-SYN compared to MP2-UNI, indicating improved performance using the synthesized WMn image. Discussion: THOMAS algorithm can successfully segment thalamic nuclei in routinely acquired bias-free MP2RAGE images with essentially equivalent quality when evaluated against WMn-MPRAGE, hence has wider applicability in studies focused on thalamic involvement in aging and disease.