Restoration of Lossy JPEG-Compressed Brain MR Images Using Cross-Domain Neural Networks

Lossy image compression allows for efficient storage and transfer of image data with varying degrees of image degradation. However, lossy compression is not commonly used in medical imaging as the process may irreversibly remove information that defines clinically important image features. The lossy component of JPEG compression is represented as lost precision in the discrete cosine transform (DCT) domain after quantization on $8 \times 8$ image blocks and results in degradation of the image. We propose a cross-domain cascade of U-nets called the W-net. This network operates in the DCT domain to restore discarded DCT coefficients that leverages information from adjacent blocks, and the image domain to suppress compression artifacts at the image pixel level. For comparison, we adapted the Automated Transform by Manifold Approximation (AUTOMAP) method for JPEG decompression by learning the dequantization of individual $8 \times 8$ DCT coefficient blocks. These results were then transformed to the image domain and processed by a U-net. The deep learning models were able to suppress common compression artifacts at the expense of high spatial frequency detail. Both the W-net and AUTOMAP network structures were quantitatively superior to standard JPEG decompression, with the W-net outperforming AUTOMAP, suggesting that leveraging DCT coefficients from adjacent blocks improves JPEG decompression performance.