Crop and Couple: cardiac image segmentation using interlinked specialist networks

TL;DR

This paper tackles automated segmentation of cardiac MRI into left ventricle, right ventricle, and myocardium, a key step in cardiovascular disease assessment. It introduces CroCNet, a two-stage pipeline where an initial E-2AUNet segments and crops the heart region, followed by weight-sharing specialist networks (LV-Net, RV-Net, MYO-Net) refined through efficient additive cross-attention to produce accurate final masks. The approach leverages an efficient additive attention mechanism to capture global context with lower complexity, and cross-attention acts as a soft relative shape prior to improve inter-anatomy consistency. On the MICCAI 2021 M&Ms-2 dataset, CroCNet achieves state-of-the-art Dice and Hausdorff distance, outperforming prior methods by notable margins, and ablations confirm the value of cropping and cross-attention for refining segmentation while maintaining single-view input efficiency.

Abstract

Diagnosis of cardiovascular disease using automated methods often relies on the critical task of cardiac image segmentation. We propose a novel strategy that performs segmentation using specialist networks that focus on a single anatomy (left ventricle, right ventricle, or myocardium). Given an input long-axis cardiac MR image, our method performs a ternary segmentation in the first stage to identify these anatomical regions, followed by cropping the original image to focus subsequent processing on the anatomical regions. The specialist networks are coupled through an attention mechanism that performs cross-attention to interlink features from different anatomies, serving as a soft relative shape prior. Central to our approach is an additive attention block (E-2A block), which is used throughout our architecture thanks to its efficiency.

Figures (5)

• Figure 1: Overview of the proposed CroCNet pipeline. (a) E-2AUNet architecture for initial segmentation in the first stage. The cropped image and initial binary segmentations are fed into (b) specialist networks (LV-Net, RV-Net, and MYO-Net) to refine the predictions for each cardiac region. We leverage (c) efficient additive attention, including (d) a cross-E2A block that implements cross-attention to intermingle features between the specialist networks.
• Figure 2: Proposed cropping process in CroCNet shows intensity and label images are cropped based on detection of the largest connected region in the inverted background prediction.
• Figure 3: Qualitative Results. Row (a) and (b) depict how the cropping process can help to remove erroneous predictions. (c) and (d) showcase the effect of utilizing the Cross-E2A attention with specialist networks.
• Figure 4: Illustration of improved segmentation of boundary regions using specialist network and Cross E-2A.
• Figure 5: Model size versus performance comparison shows average Dice score and number of model parameters.