Deep Reinforcement Learning for Vessel Centerline Tracing in Multi-modality 3D Volumes

Jan 2019

tl;dr: Use DRL for vessle centerline tracing

Overall impression

The paper proposes a new method for tracing using deep reinforcement learning (DRL). However the paper uses a heavily handcrafted reward function which may be simplified.


Key ideas

\(D(p_t, G) = ||\lambda(p_t -g_{d+k}) + (1-\lambda) (g_{d+k+1} - g_{d+k-1})||\)

\(r_t = D(p_t, G) - D(p_{t+1}, G), \text{ if } ||p_t - g_d|| <=l; \\ r_t = ||p_t - g_d|| - ||p_{t+1} - g_d||, \text{ otherwise.}\)

Technical details