Learning high-order spatial interactions and cascaded attention decoding for MRI-based ventricular segmentation

Abstract

The precise segmentation of ventricles facilitates the provision of quantitative information concerning cardiac structure and function to medical professionals, thereby enabling accurate diagnosis. The present challenges in ventricular segmentation are twofold: Capturing positional information and morphological features of the ventricular region is challenging due to the cyclic beating of the heart, and ventricles exhibit significant similarity to other tissues in greyscale values. Compared to Convolution neural networks (CNNs) that lack the ability to establish long-range dependencies, Transformer which could perceive global contextual information have made significant progress in ventricular segmentation. However, Vanilla Transformer neglects cross-channel dependencies that can be utilized to capture positional and morphological features of ventricles hidden in numerous channels. To address this problem, this work proposes Efficient Split Channel Product Net (ESCP-Net). Specifically, ESCP-Net designs a Split-Channel Product Encoding Module (SCPEM) to grasp the ventricular position and morphological information by means of High Order Spatial interactions. Moreover, ESCP-Net cascades the Non-isomorphic Feature Fusion Module (NFFM) and the attention mechanism, thereby fusing the hierarchical feature representations and enlarging the feature grey scale difference to generating clearer feature representation. To verify this proposed method, extensive experiments on three heart chambers segmentation datasets (Automated Cardiac Diagnosis Challenge[ACDC]2017, Right ventricle Segmentation Challenge[RVSC]2012, Challenge of Left Ventricular Segmentation in Heart MR[SunnyBrook]2009) are conducted. The proposed method achieved the best performance with a DSC of 91.66% in ACDC datasets and 85.54% in RVSC and 95.94% in SunnyBrook. This work demonstrated that our method provides a promising solution for ventricular Segmentation.