Steps Toward the Automatic Labeling of 3D Medical Images

Applications of Computer Vision in Medical Image Processing

Volume

Papers from the 1994 AAAI Spring Symposium

Contents

We summarize here our latest works about the extraction of stable feature lines and points from 3D medical images, and their use in automated protocols. Mainly, we have shown that it is possible to extract automatically, with sub-voxel accuracy, specific lines and points from 3D images, based on differential geometry criteria. These are the crest lines (or eztremal lines), corresponding to the loci of the surface whose maximal curvatures in absolute value are locally extremal, and also the eztremal points, which are specific points of the extremal lines, also invariant to rigid transforms. Recently, we have shown that it is possible to extract an invariant graph from the 3D surface, the eztremal mesh, based on a new local "invariant that we call the Gav.~sian eatremality. The vertices of the extremal mesh are the extremal points and the edges are the extremal lines (see [Thirion, 1993s]). We have used successfully those features to perform the automatic registration in the mono-modal, mono-patient case.