Research map / Image generation / Image restoration
Orochi: Versatile Biomedical Image Processor
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.
Introduces Orochi, a pre-trained versatile biomedical image processor (Multi-head Hierarchy Mamba) trained with task-related joint-embedding pre-training and fine-tuned via Full/Normal/Light tiers for registration, fusion, restoration and super-resolution.
Problem
Specialist biomedical low-level image processing models are task- and dataset-specific, inefficient and impractical for biologists who need an all-in-one workflow across registration, fusion, restoration and super-resolution.
Approach
Pre-train a single hierarchical Mamba encoder on raw data from 100+ public studies using Random Multi-scale Sampling and self-supervision based on biomedical degradations (dual masking, spatially-varying Gaussian downsampling, Perlin-noise deformation, multi-stage noise) instead of MIM, then adapt with three tiers of fine-tuning including depth-wise separable convolution heads for parameter-effici
Key results
- On isotropic 3D restoration (CARE), Orochi (Light) reaches 29.77 XY PSNR vs 27.67 for UniFMIR (27.12 for UniFMIR's XY), and Orochi (Full) is +0.85 PSNR over UniFMIR.
- On the inter-patient brain registration task (OASIS), Orochi (Full) achieves 83.62 Dice vs 82.22 for TransMorph-L and 81.81 for MambaMorph.
- On CT-MRI fusion (VIFB), Orochi (Full) reaches Qabf 0.41 vs 0.39 for BSAFusion, with Qcv 2,351.57 vs 4,155.10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TJP pre-training instead of MAE single-mask pre-training (ablation) | Orochi · Registration | Dice: 71.22 -> 83.62 | Helped |
| Use TJP pre-training instead of I-JEPA dual-mask pre-training (ablation) | Orochi · Registration | Dice: 69.97 -> 83.62 | Helped |
| Use TJP pre-training instead of MAE single-mask pre-training (ablation) | Orochi · Fusion | Qabf: 0.36 -> 0.41 | Helped |
| Use TJP pre-training instead of I-JEPA dual-mask pre-training (ablation) | Orochi · Fusion | Qabf: 0.39 -> 0.41 | Helped |
| Use TJP pre-training instead of MAE single-mask pre-training (ablation) | Orochi · Restoration | PSNR: 26.67 -> 29.88 | Helped |
| Use TJP pre-training instead of I-JEPA dual-mask pre-training (ablation) | Orochi · Restoration | PSNR: 25.02 -> 29.88 | Helped |
| Use TJP pre-training instead of MAE single-mask pre-training (ablation) | Orochi · Super-resolution | PSNR: 29.17 -> 33.63 | Helped |
| Use TJP pre-training instead of I-JEPA dual-mask pre-training (ablation) | Orochi · Super-resolution | PSNR: 28.81 -> 33.63 | Helped |
| Fine-tune with the Light depth-wise separable convolution head instead of Full fine-tuning (ablation) | Orochi · CARE, XY | PSNR: 28.31 -> 29.77 | Helped |
| Fine-tune with the Light head instead of Full fine-tuning (ablation) | Orochi · OASIS | Dice: 83.62 -> 79.61 | Hurt |
| Fine-tune with the Light head instead of Full fine-tuning (ablation) | Orochi · HBA, 4mm | PSNR: 35.33 -> 34.83 | Hurt |
| Fine-tune with the Light head instead of Full fine-tuning (ablation) | Orochi · VIFB | Qabf: 0.41 -> 0.34 | Hurt |
Surprising findings
- Parameter-efficient Light tuning outperformed full fine-tuning on the isotropic restoration task, attributed to overfitting on fewer than 100 training patches.
- Masked image modelling (MAE) gives poor registration performance (71.22 Dice vs 83.62 for TJP), and I-JEPA dual masking underperforms the proposed TJP.
Limitations
- Scaling law of the method during pre-training was not investigated due to compute limits.
- Focus restricted to low-level tasks; no high-level understanding tasks evaluated.
- Some fusion baselines show limited registration capability, so cross-task comparison is not like-for-like.
Experimental setup
Datasets: raw data from over 100 public biomedical studies CARE HBA OASIS VIFB IXI BioSR
Benchmarks: CARE HBA OASIS (Learn2Reg 2021) VIFB IXI BioSR
Metrics: PSNR SSIM Dice HD95 SDlogJ Qabf Qcv
Baselines: UniFMIR SwinIR MambaIR CARE InverseSR VCM LIIF TransMorph MambaMorph BSAFusion MAE I-JEPA
Builds on: JEPA I-JEPA Mamba Swin Transformer UniFMIR
Closest papers by meaning
- BIOMEDICA: An Open Biomedical Image-Caption Archive, Dataset, and Vision-Language Models Derived from Scientific Literature (CVPR 2025 · similarity 0.82)
- LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding (CVPR 2026 · similarity 0.82)
- Specialized Foundation Models Struggle to Beat Supervised Baselines (ICLR 2025 · similarity 0.81)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.81)
- Fish-Vista: A Multi-Purpose Dataset for Understanding & Identification of Traits from Images (CVPR 2025 · similarity 0.81)
- Revisiting Convolution Architecture in the Realm of DNA Foundation Models (ICLR 2025 · 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.