3D Medical Image Segmentation

This field develops methods that label anatomical structures and disease in 3D CT, MRI, PET, and other scans, supporting diagnosis, treatment planning, and quantitative analysis. U-Net and nnU-Net remain common baselines, alongside Transformer and state-space models; newer work also uses anatomical and topological constraints, foundation models with prompts, and efficient ways to process volumes. To reduce dependence on costly voxel-level labels, researchers use self-supervised pretraining, synthetic data, semi-supervised learning, and interactive correction, evaluating methods on tasks such as BraTS, BTCV, and ACDC. A central challenge is that strong scores on familiar datasets may not hold across hospitals, scanners, or modalities, while thin structures, uncertain boundaries, and limited compute make reliable clinical deployment difficult.