LFC-UNet: learned lossless medical image fast compression with U-Net

Area of Science:

• Medical Imaging
• Computer Vision
• Data Compression

Background:

• Rapid advancements in medical technology necessitate efficient medical image compression.
• Neural networks offer promising solutions for lossless compression, but existing methods struggle with residual data imbalance.
• Accurate residual probability estimation is crucial for effective learning-based compression.

Purpose of the Study:

• To develop an efficient and accurate lossless compression method for medical images using neural networks.
• To address the challenge of excessive residual concentration in existing compression techniques.
• To improve the speed and performance of neural network-based medical image compression.

Main Methods:

• Employed a weighted cross-entropy method to manage imbalanced residual categories.
• Integrated U-Net architecture with skip connections for enhanced feature capture and probability estimation.
• Developed a framework enabling single-pass inference for residuals and their probabilities.

Main Results:

• Achieved state-of-the-art performance on medical datasets for lossless image compression.
• Demonstrated significantly faster processing speeds compared to existing methods.
• An example using head CT data showed 2.30 bits per pixel compression efficiency in 0.320 seconds.

Conclusions:

• The proposed weighted cross-entropy and U-Net-based approach effectively overcomes residual data imbalance in neural network compression.
• This method offers superior compression efficiency and processing speed for medical images.
• The framework's single-pass inference capability contributes to its remarkable encoding speed.