Tempera: Spatial Transformer Feature Pyramid Network for Cardiac MRI Segmentation

Assessing the structure and function of right ventricle (RV) is important in diagnosis several cardiac pathologies. However, it remains more challenging to segment RV than left (LV). In this paper, we focus on segmenting both short (SA) long-axis (LA) MR images simultaneously. For task, propose a new multi-input/output architecture, hybrid 2D/3D geometric spatial TransformEr Multi-Pass fEature pyRAmid (Tempera). Our feature pyramid extends current designs by allowing not only multi-scale output but SA LA input as well. Tempera transfers learned features between via layer weight sharing incorporates target transformer map predicted segmentation space. model achieves an average Dice score 0.836 0.798 for LA, respectively, 26.31 mm 31.19 Hausdorff distances. This opens up potential incorporation models into clinical workflows.