Research map / Image generation / Image restoration
WISE: A Framework for Gigapixel Whole-Slide-Image Lossless Compression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WISE, a non-neural, dictionary-based lossless compressor for gigapixel whole-slide images that combines hierarchical delta projection, bitmap regrouping, and LZW coding.
Problem
Lossless compression of gigapixel whole-slide images (WSIs) is poorly handled by existing compressors (PNG, Huffman, arithmetic, Gzip, Zstd, LZMA, neural codecs) because WSIs have high-frequency, irregular information patterns.
Approach
Reduce entropy with a hierarchical first-order delta projection across row, column, and channel, then regroup bits by position across bytes (bitmap coding) so that repeated patterns emerge, and finally apply LZW dictionary coding to exploit them.
Key results
- WISE compresses gigapixel WSIs to 36x on average and up to 136x (C17 sample image 5).
- On non-empty patches, WISE reaches 2.82 (C16), 3.20 (C17), 2.50 (TCGA) vs. ArIB-BPS 2.63, 2.72, 2.38.
- Ablation on C16: LZMA baseline 1.25 rises to 2.12 with hierarchical linear projection and 2.82 with bitmap encoding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use WISE instead of ArIB-BPS | WSI patch compression · C16 patches | compression ratio: 2.63 -> 2.82 | Helped |
| use WISE instead of ArIB-BPS | WSI patch compression · C17 patches | compression ratio: 2.72 -> 3.20 | Helped |
| use WISE instead of ArIB-BPS | WSI patch compression · TCGA patches | compression ratio: 2.38 -> 2.50 | Helped |
| use WISE instead of Lzma | WSI patch compression · BNCB patches | compression ratio: 1.29 -> 1.63 | Helped |
| use WISE instead of ArIB-BPS | WSI patch compression · IHC4BC patches | compression ratio: 3.81 -> 3.85 | Helped |
| add hierarchical linear projection to the LZMA baseline (ablation) | WISE · C16 patches | compression ratio: 1.25 -> 2.12 | Helped |
| add bitmap encoding after hierarchical linear projection (ablation) | WISE · C16 patches | compression ratio: 2.12 -> 2.82 | Helped |
| add hierarchical linear projection to the LZMA baseline (ablation) | WISE · C17 patches | compression ratio: 1.39 -> 2.46 | Helped |
| add bitmap encoding after hierarchical linear projection (ablation) | WISE · C17 patches | compression ratio: 2.46 -> 3.20 | Helped |
| add hierarchical linear projection to the LZMA baseline (ablation) | WISE · TCGA patches | compression ratio: 1.17 -> 2.03 | Helped |
| add bitmap encoding after hierarchical linear projection (ablation) | WISE · TCGA patches | compression ratio: 2.03 -> 2.50 | Helped |
Surprising findings
- Standard neural and entropy-based compressors such as PNG and Huffman achieve little to no compression on WSI images, while LZMA performs comparatively well.
- Dictionary-based methods improve compression as input size grows, unlike entropy-based methods.
Limitations
- Compresses only the base pyramid level and focuses on non-empty informative regions; empty-area search is out of scope.
- Ignores the alpha channel, which may discard some information.
- Throughput and memory grow with input size due to LZW dictionary construction.
Experimental setup
Datasets: C16 C17 TCGA BNCB DigestPath IHC4BC
Benchmarks: C16 C17 TCGA BNCB IHC4BC
Metrics: compression ratio throughput memory usage PSNR (bitmap similarity)
Baselines: PNG TIFF Huffman Arithmetic coding Gzip Zstd-22 LZMA ArIB-BPS Pfor PAC
Builds on: LZW LZMA Zstandard ArIB-BPS
Closest papers by meaning
- Context Matters: Query-aware Dynamic Long Sequence Modeling of Gigapixel Images (ICML 2025 · similarity 0.82)
- Large Language Models for Lossless Image Compression: Next-Pixel Prediction in Language Space is All You Need (NeurIPS 2025 · similarity 0.82)
- Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach (ICML 2024 · similarity 0.81)
- Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation (NeurIPS 2026 · similarity 0.81)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.81)
- Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.