Reproduce “A Flexible 2.5D Medical Image Segmentation Approach with In-Slice and Cross-Slice Attention” on OpenResearch

About this paper

Amarjeet Kumar, Hongxu Jiang, Muhammad Imran, Cyndi Valdes, Gabriela Leon, Dahyun Kang, Parvathi Nataraj, Yuyin Zhou and 2 more

CSA-Net, a 2.5D medical image segmentation model, utilizes in-slice and cross-slice attention mechanisms alongside a Vision Transformer to integrate 3D spatial context while processing images with varying numbers of slices. The model achieved improved Dice coefficients and reduced Hausdorff distances across brain, prostate capsule, and multi-class prostate zone MRI segmentation tasks, outperforming existing 2D and 2.5D methods.

Key insights

• Effectively integrating both in-slice and cross-slice pixel-level attention allows for a comprehensive understanding of 3D spatial context from anisotropic medical volumes within an efficient 2.5D framework.
• Using a Vision Transformer encoder on aggregated attention features helps capture global contextual information, enhancing segmentation accuracy, especially for complex or small structures.
• Designing a 2.5D architecture that can adapt to an arbitrary number of slices by focusing on local 3-slice contexts ensures flexibility and applicability across diverse clinical datasets with varying through-plane resolutions.

Summary by alphaXiv.