Parallel Hyper-Prior and Dual-Domain Autoregressive Transformer for Lossless JPEG Recompression of Pathology Images

Area of Science:

• Medical Imaging
• Computer Vision
• Data Compression

Background:

• Pathology images require significant storage due to high resolution.
• Existing lossy compression (JPEG) still results in considerable storage needs.
• Lossless recompression of already lossy-compressed JPEG images is an unexplored challenge.

Purpose of the Study:

• To develop an effective method for lossless recompression of JPEG-compressed pathology images.
• To reduce storage costs associated with high-resolution medical imaging.
• To improve compression performance beyond existing techniques.

Main Methods:

• Proposed a Parallel Hyper-Prior and Dual-Domain Autoregressive Transformer model.
• Processed discrete cosine transform (DCT) components in parallel to reduce attention complexity.
• Utilized spatial fractal and frequency-domain autoregressive transformers for hierarchical and frequency-based dependencies.
• Introduced a mixture entropy model for accurate entropy estimation and enhanced conditional modeling.

Main Results:

• Achieved state-of-the-art performance in lossless JPEG recompression.
• Attained up to 33.4% additional lossless compression on standard JPEG files.
• Demonstrated strong generalization capabilities on out-of-distribution datasets (TCGA).

Conclusions:

• The proposed model significantly enhances lossless compression for JPEG-compressed pathology images.
• The method offers substantial storage savings for medical imaging data.
• The approach shows promise for real-world applications requiring efficient medical image storage and retrieval.