Multi-view Cardiac Image Segmentation via Trans-Dimensional Priors

Our Contributions

• We propose a sequential 3D-to-2D-to-3D approach for multi-view cardiac image segmentation by effectively utilizing the trans-dimensional segmentation priors (TDSP), which transform a segmentation from one view into another and serve as guidance.
• The TDSP provides a robust anatomical reference at the network’s input and encourages the network to produce anatomically plausible segmentation maps.
• We also introduce a Heart Localization and Cropping (HLC) module to focus the segmentation on the heart region only. This strategy reduces the computation for the second and third-stage segmentation network and eliminates false positive predictions.
• Extensive experiments are conducted to showcase the efficacy of the proposed pipeline utilizing the HLC module and TDSP, where our proposed method outperforms the state-of-the-art as well as methods on the M&Ms-2 challenge leaderboard.

  Paper Available Here