3D CNNs

3D CNNs apply convolution across three dimensions to capture structure in medical scans, video, satellite time series, and scientific simulations. Common approaches include encoder-decoder networks such as 3D U-Nets, residual networks, and spatiotemporal models, with transfer learning, self-supervision, and knowledge distillation used to improve learning or reduce model size. These models can use context that 2D methods miss, but their greater computational cost makes lightweight convolutions, pruning, and efficient training important. A central tension is whether full 3D context reliably improves performance: 2D or 2.5D inputs can be more efficient or stable, so careful subject-level, cross-dataset, and resource-aware evaluation remains essential.