Deep Learning Based Rib Centerline Extraction and Labeling

Feb 2019

tl;dr: Locate the first and last pairs of ribs and use as anchor for labeling

Overall impression

The paper proposes a smart idea of classifying all 12 pairs of ribs into three classes: first, intermediate and last. The final results of ~80% success rate is a bit lacking though. Previous methods of detecting seed points and track has a disadvange that it often misses an entire rib if the seed is missing. This paper generates rib candidates first and then traced to generate centerline.

Key ideas

Technical details