Content-Aware Image Compression Model for Macromolecular Crystallography Data

The structural biology community faces the challenge of managing massive data volumes from macromolecular crystallography (MX) experiments, motivating the development of efficient lossy compression methods for permanent data archival. In this paper, we introduce a novel content-aware deep image compression model specifically designed for MX data. Bragg reflections (also called spots) position information derived from a spot finder is incorporated during training to guide the compression preserving content-relevant regions, while inference remains fully independent of Bragg spot information. We further investigate different mask generation strategies to balance the trade-off between compression ratio and diffraction pattern preservation. Additionally, we propose novel content-aware loss functions to better preserve Bragg spot information. Experimental results demonstrate that our method achieves both higher spot detection performance and superior compression ratio compared to other methods, enabling scientifically reliable compression for crystallographic data archiving and analysis. Our code is available at https://github.com/DJX1995/BNL-ImageCompression.