APRIL: Anatomical prior-guided reinforcement learning for accurate carotid lumen diameter and intima-media thickness measurement.

摘要: Carotid artery lumen diameter (CALD) and carotid intima-media thickness (CIMT) are essential factors for estimating the risk of many cardiovascular diseases. The automatic measurement them in ultrasound (US) images is an efficient assisting diagnostic procedure. Despite advances, existing methods still suffer issue low measuring accuracy poor prediction stability, mainly due to following disadvantages: (1) ignore anatomical prior prone give anatomically inaccurate estimation; (2) require carefully designed post-processing, which may introduce more estimation errors; (3) rely on massive pixel-wise annotations during training; (4) can not estimate uncertainty predictions. In this study, we propose Anatomical Prior-guided ReInforcement Learning model (APRIL), innovatively formulate CALD & CIMT as RL problem dynamically incorporate (AP) into system through a novel reward. With guidance AP, keypoints APRIL avoid various anatomy impossible mis-locations, accurately measure based their corresponding locations. Moreover, formulation significantly reduces human annotation effort by only using several help eliminate extra post-processing steps. Further, module variance, guide us adaptively rectify those frames with considerable uncertainty. Experiments challenging US dataset show that achieve MAE (in pixel/mm) 3.02±2.23 / 0.18±0.13 CALD, 0.96±0.70 0.06±0.04 CIMT, surpass popular approaches use annotations.