Crop and Couple: Cardiac Image Segmentation using interlinked specialist networks

Our Contributions

• We propose CroCNet, a novel two-stage architecture that computes a first segmentation used to identify anatomies and perform cropping on the original image. The cropped data is fed into a second stage consisting of coupled specialist networks that perform binary segmentation.
• For the first stage of CroCNet, we propose E-2AUNet, a novel hybrid encoder-decoder architecture that modifies a UNet with E-2A blocks.
• In the second stage, we implement specialist networks coupled through efficient additive cross-attention, which acts as a soft shape prior.
• Our results set a new state-of-the-art on the M&Ms-2 dataset, outperforming the recently proposed TransFusion.

  Paper Available Here