Conditional random fields for tubulin-microtubule segmentation in cryo-electron tomography

Cryo-electron tomography allows 3D observation of biological specimens in their native and hydrated state at high spatial resolution (4-5 nanometers). Traditionally cryo-tomograms have very low signal-to-noise ratios and conventional image segmentation methods are limited yet. In this paper, we formulate the segmentation problem of both small tubulin aggregates and microtubules against the background as a two class labeling problem in the Conditional Random Field framework. In our approach, we exploit image patches to take into account spatial contexts and to improve robustness to noise. Because of the contrast anisotropy in the specimen thickness direction, each 2D section of the 3D tomogram is segmented separately with an optional update of reference patches. This method is evaluated on synthetic data and on cryo-electron tomograms of in vitro microtubules.