Model Error Maps as Robust Features for T1/T2-Weighted Data Segmentation

Umesh Suryanarayana Rudrapatna¹, Annette van der Toorn¹, Ivo A. Tiebosch¹, Josien P.W Pluim¹, Rick M. Dijkhuizen¹

T1- and T2-weighted data segmentation has hitherto relied upon parametric maps and intensities as primary features. The parametric maps contain the necessary information explicitly, but become unreliable at low SNR. To overcome this hurdle, building on estimation theory results, we propose using functions based on fitting-errors when specific relaxivities are assumed. This approach leads us to a simpler linear estimation problem and provides for incorporation of prior knowledge about relaxivities, the reliability of which does not critically affect the outcome. Feature selection filters, the results of which generalize to a broad class of supervised and unsupervised learning scenarios testify the merits of this strategy.