🤖 AI Summary
To address the challenges of automatic segmentation in late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) images—namely, indistinct infarct scar boundaries, low contrast, and artifact interference—this paper proposes a 2D-3D cascaded CNN framework. First, a U-Net–style 2D encoder-decoder generates an initial segmentation; then, a lightweight 3D ResCNN refinement module models volumetric contextual information via joint channel-spatial attention. A novel residual-guided feedback mechanism explicitly models and iteratively corrects cross-dimensional feature discrepancies between the 2D and 3D modules. Additionally, an uncertainty-weighted loss function enhances segmentation robustness. Evaluated on the MyoPS 2020 dataset, the method achieves a Dice score of 89.7%, outperforming a single-stage 3D U-Net by 4.2 percentage points, while significantly improving recall for small lesions and boundary localization accuracy.
