TCAM-Diff generates high-res 3D medical images with less memory

TCAM-Diff (Triplane-Aware Cross-Attention Medical Diffusion Model) tackles a key bottleneck in 3D medical imaging: the massive memory needed to encode and generate high-resolution volumetric data. The model first uses a decoder-only autoencoder to learn a compact triplane representation from dense volumes—three 2D feature planes that encode 3D information efficiently. Generalization operations prevent overfitting during this compression. Then a triplane-aware cross-attention diffusion model learns and integrates these features, and a pre-trained decoder quickly transforms generated features back into full 3D volumes.

Experiments on three public datasets—BrainTumour 128×128×128, Pancreas 256×256×256, and Colon 512×512×512—show TCAM-Diff outperforming existing encoder-decoder methods with similar latent space sizes. The authors used MSE and SSIM for reconstruction quality and a Wasserstein GAN critic for generative quality. The code is open-source, and the paper has been accepted at AAAI 2025, signaling strong peer validation. This work enables memory-efficient generation of high-resolution medical images, potentially improving training data for diagnostic AI and reducing hardware barriers in medical research.