Lossless medical image compression based on anatomical information and deep neural networks

Modern imaging modalities generate large volumes of medical data that place a heavy burden on both storage and transmission. Consequently, image data compression is a key research topic in the field of medical imaging. This paper proposes a compression technique that combines anatomical information and a custom deep neural network model to facilitate the transmission and storage of medical image data. This technique first divides the medical image data into specific regions based on anatomical features, namely, image density, relative position, and organ size. A deep neural network is then trained to generate a series of optimal predictors for each region. These predictors can be adaptively switched based on the characteristics of the region being compressed. The residuals are finally compressed with an entropy coding scheme. An evaluation of this compression technique shows that the proposed divide and conquer method indeed achieves high prediction accuracy and obtains a compression performance that is better than that of JPEG2000 by 38%.

Automated summary

This paper introduces a medical image data compression technique that combines anatomical information and a custom deep neural network model to improve the efficiency of data transmission and storage. Through dividing specific regions and training optimized predictors, high prediction accuracy and better compression performance than JPEG2000 have been achieved.