Researchers unveil AI-free lossless 3D medical image compression
New tri-plane context model beats deep learning methods without training data
Lossless compression of volumetric medical images is critical when data fidelity is non-negotiable, yet traditional methods are held back by rigid handcrafted models and deep learning approaches demand heavy compute. Now, researchers introduce a tri-plane context tree (TCT) method that delivers performance on par with recent DNN-based techniques—without using neural networks or external training data. By decomposing complex 3D context modeling into efficient 2D modeling across three orthogonal planes, the method captures both intra-slice and inter-slice redundancies. An adaptive binary tree structure is built specifically for each input volume, selecting from tri-plane-based predictors and feature extractors at every node, so context modeling adapts to local structures. The model is learned directly from a sampled subset of the volume using minimum description length optimization, then each pixel retrieves its context, computes a residual, and entropy-codes it with corresponding histograms. The result: strong compression across multiple datasets, low computational cost, and fast coding speeds—making it highly practical for real-world deployment.
- Tri-plane Context Tree (TCT) achieves DNN-level compression without neural networks or training data
- Model dynamically adapts to local structures using adaptive binary trees and tri-plane predictors
- Tested across multiple datasets with compression speeds suitable for resource-constrained medical environments
Why It Matters
Enables high-fidelity medical image storage and transmission in hospitals with limited computational resources