Opportunity
The advancement of medical imaging technologies, such as MRI and CT, has led to the widespread use of 3D medical images in diagnosis and treatment. However, these images suffer from dramatically increased data volume compared to 2D images, making efficient lossless compression crucial for storage and transmission. Traditional methods rely on hand-crafted components like 3D transforms (e.g., DCT, DWT) or sequence-based techniques, which lack adaptability to the specific characteristics of 3D medical images. Deep learning-based methods have shown promise but often fail to fully exploit the inherent bilateral correlations (e.g., anatomical symmetry and inter-slice redundancies) in 3D medical images. This patent addresses these limitations by proposing a novel lossy-then-lossless compression framework that leverages bilateral context modeling to achieve superior compression performance.
Technology
The patent introduces a hybrid compression framework combining lossy and lossless phases. First, the 3D medical image is partitioned into 2D slices, which are encoded using a lossy codec (e.g., VVC, HEVC) to produce compact bitstreams. The lossy reconstructions are then compared to the original slices to compute residuals. These residuals, containing intricate details not captured by the lossy compression, are encoded losslessly using a Bilateral Context Modeling Network (BCM-Net). The BCM-Net exploits intra-slice correlations (via a Symmetry-based Intra-slice Context Extraction module) and inter-slice correlations (via a Bi-directional Inter-slice Context Extraction module) to generate precise probability distributions for arithmetic coding. The intra-slice module leverages anatomical symmetry, while the inter-slice module uses bi-directional references from a hierarchical-B coding structure. This approach ensures efficient redundancy reduction and outperforms state-of-the-art methods.
Advantages
- High Compression Efficiency: Achieves better compression ratios than traditional and learned methods by exploiting bilateral correlations.
- Adaptability: Compatible with existing 3D medical image pipelines and standard codecs (e.g., VVC).
- Flexible Output: Provides both lossy reconstructions for quick review and lossless reconstructions for precise diagnosis.
- Computational Feasibility: Reduced encoding complexity compared to methods requiring online optimization.
Applications
- Medical Imaging: Compression of MRI, CT, and other volumetric medical images for storage and transmission.
- Telemedicine: Enables efficient sharing of high-resolution 3D medical images over networks.
- Archival Systems: Reduces storage costs for large-scale medical image databases.
- Real-Time Processing: Supports rapid access and processing of compressed 3D images in clinical settings.
