Automated olfactory bulb segmentation on high resolutional T2-weighted MRI

• DOI: 10.1016/j.neuroimage.2021.118464
• Publication Date: 2021-08-10T07:18:42Z
• Authors: Santiago Estrada; Ran Lu; Kersten Diers; Weiyi Zeng; Philipp Ehses; Tony Stöcker; Monique M. B Bret...; Martin Reuter
• Keywords: Hausdorff distance; Similarity (geometry); Visibility; Dice

The neuroimage analysis community has neglected the automated segmentation of olfactory bulb (OB) despite its crucial role in function. lack an automatic processing method for OB can be explained by challenging properties (small size, location, and poor visibility on traditional MRI scans). Nonetheless, recent advances acquisition techniques resolution have allowed raters to generate more reliable manual annotations. Furthermore, high accuracy deep learning methods solving semantic problems provides us with option reliably assess even small structures. In this work, we introduce a novel, fast, fully pipeline accurately segment tissue sub-millimeter T2-weighted (T2w) whole-brain MR images. To end, designed three-stage pipeline: (1) Localization region containing both OBs using FastSurferCNN, (2) Segmentation within localized through four independent AttFastSurferCNN - novel architecture self-attention mechanism improve modeling contextual information, (3) Ensemble predicted label maps. For were manually annotated total 620 T2w images training (n=357) testing. exhibits performance terms boundary delineation, localization, volume estimation across wide range ages 203 participants Rhineland Study (Dice Score (Dice): 0.852, Volume Similarity (VS): 0.910, Average Hausdorff Distance (AVD): 0.215 mm). Moreover, it also generalizes scans dataset never encountered during training, Human Connectome Project (HCP), different parameters demographics, evaluated 30 cases at native 0.7 mm HCP (Dice: 0.738, VS: 0.790, AVD: 0.340 mm), default 0.8 0.782, 0.858, 0.268 We extensively validated our not only respect but known effects, where sensitively replicate age effects (β=-0.232, p<.01). analyze 3D less than minute (GPU) end-to-end fashion, providing validated, efficient, scalable solution automatically assessing volumes.