Practical Lossless Volumetric Medical Image Compression via Tri-Plane Context Tree Learning.

Area of Science:

• Medical Imaging
• Data Compression
• Computer Vision

Background:

• Lossless compression is crucial for volumetric medical images in clinical and research settings.
• Traditional methods lack efficiency, while deep neural network (DNN) methods require significant computational resources.
• Resource-constrained environments face challenges deploying advanced compression techniques.

Purpose of the Study:

• To develop a high-performance, lossless compression method for volumetric medical images.
• To create a method that avoids deep neural networks (DNNs) and external training data.
• To address the limitations of existing compression techniques in terms of efficiency and computational cost.

Main Methods:

• Proposed a novel tri-plane context tree (TCT)-based method for lossless volumetric medical image compression.
• Introduced a compact tri-plane context representation for efficient 3D context modeling.
• Developed an input-specific TCT model with adaptive binary tree structure, dynamically selecting predictors and feature extractors.
• Learned the TCT model by optimizing minimum description length (MDL) from a subset of the input volume, avoiding offline training.

Main Results:

• Achieved compression performance comparable to recent deep neural network (DNN)-based methods across multiple datasets.
• Demonstrated low computational cost and fast coding speeds.
• The method is highly applicable in practical, resource-constrained settings.

Conclusions:

• The proposed TCT-based method offers a viable alternative for lossless volumetric medical image compression.
• It provides high compression efficiency without the computational burden of DNNs.
• The method's adaptability and efficiency make it suitable for real-world clinical and research applications.