LC-TMNet: learned lossless medical image compression with tunable multi-scale network

Area of Science:

• Medical Imaging
• Computer Science
• Data Compression

Background:

• High-quality medical images are vital for accurate diagnosis, necessitating lossless compression to maintain image integrity.
• Neural networks integrated with entropy encoders offer advanced lossless compression, outperforming traditional methods.
• Existing neural network compression methods struggle with probabilistic estimation in complex or edge regions, limiting performance.

Purpose of the Study:

• To develop a novel lossless image compression method that overcomes limitations in probabilistic estimation for complex image regions.
• To improve the accuracy and efficiency of neural network-based lossless image compression for medical applications.

Main Methods:

• A flexible tree-structured image segmentation mechanism was employed to leverage relationships between subimages.
• An attention mechanism was integrated into the UNet network architecture to enhance probabilistic estimation accuracy.
• Variable-speed compression was implemented, offering distinct fast and slow modes.

Main Results:

• The proposed method demonstrates improved probabilistic estimation, particularly in complex textured regions.
• Variable-speed compression achieved state-of-the-art compression speed in fast mode.
• The slow mode achieved state-of-the-art compression performance.

Conclusions:

• The novel method effectively enhances lossless image compression accuracy and efficiency for medical imaging.
• Flexible tree-structured segmentation and attention mechanisms are key to improving neural network-based compression.
• The variable-speed compression offers practical adaptability for different clinical or research needs.