Echoquan-Net: Direct Quantification Of Echo Sequence For Left Ventricle Multidimensional Indices Via Global-Local Learning, Geometric Adjustment And Multi-Target Relation Learning

Rongjun Ge, Guanyu Yang, Chenchu Xu, Jiulou Zhang, Yang Chen, Limin Luo, Cheng Feng, Heye Zhang, Shuo Li

DOI: https://doi.org/10.1007/978-3-030-30493-5_24

2019-01-01

Abstract:Accurately quantifying multidimensional indices of the left ventricle (LV) in 2D echocardiography (echo) is clinically significant for cardiac disease diagnosis. However, the challenges of the frequently missing information, the high geometric variability and the uncertain multi-dimensional indices relation hinder its automated analysis development. Here, we propose an EchoQuan-Net to directly quantify LV in echo sequence from the multidimension, covering length and width for 1D, area for 2D, and volume for 3D. The net consist of three components: (1) Global-Local Learning to capture contextual information in the cardiac cycle for each frame, with global information from whole sequence and local information from the individual frame; (2) Geometric Adjustment to promote a canonical region of interest for LV, with translation, rotation and scale invariant; (3) Multi-target Relation learning to promote joint quantification for LV multidimensional indices, with sparse latent regression. The experiments reveal that EchoQuan-Net gains high accuracy, with mean accuracy error of 3.14 mm, 3.10 mm, 276mm(2) and 13.5ml for length, width, area and volume. The results show great potential of our method in clinical cardiac function assessment.