Research map / Image generation / Autoregressive and latent generation
Unraveling Normal Anatomy via Fluid-Driven Anomaly Randomization
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 UNA, a modality-agnostic 3D model that reconstructs healthy brain anatomy from MRI/CT scans with pathology, trained on synthetic fluid-driven (advection-diffusion) anomaly randomization plus real healthy and stroke data.
Problem
Contrast- and resolution-agnostic brain image analysis models assume healthy anatomy and degrade on scans with extensive pathology, while existing pathology-aware methods need paired pathology annotations and fine-tuning.
Approach
Generate unlimited realistic pathology profiles on-the-fly by forward advection-diffusion PDEs (incompressible flow and non-negative diffusion from Perlin-noise potentials) seeded with a few real lesion maps or random shapes, encode them into synthetic healthy scans of random modality, and train with contralateral-paired input and an intra-subject self-contrastive loss.
Key results
- On simulated T1w MRI with pathology, UNA reaches 31.98 PSNR / 0.981 SSIM full brain versus 22.86 / 0.859 for Brain-ID, and 25.63 PSNR on CT versus 20.15 for Brain-ID.
- On real stroke data (ATLAS), UNA reaches 29.10 PSNR / 0.974 SSIM on healthy regions, outperforming PEPSI (26.73 / 0.905).
- For anomaly detection by reconstruction difference, UNA attains a mean Dice of 0.31-0.37 across datasets, above Brain-ID (0.24-0.28), VAE and LDM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UNA instead of Brain-ID | healthy anatomy reconstruction · Simulated pathology, T1w MRI, full brain | L1: 0.0231 -> 0.0147 | Helped |
| use UNA instead of Brain-ID | healthy anatomy reconstruction · Simulated pathology, T1w MRI, diseased region | PSNR: 40.22 -> 45.61 | Helped |
| use UNA instead of SynthSR | anomaly detection · Simulated pathology, ADNI | Dice: 0.27 -> 0.36 | Helped |
| use UNA instead of SynthSR | anomaly detection · Stroke T1w MRI, ATLAS | Dice: 0.24 -> 0.31 | Helped |
| add fluid-driven anomaly randomization to training (ablation) | UNA · Simulated pathology, T1w MRI, full brain | L1: 0.0229 -> 0.0147 | Helped |
| initialize fluid-driven anomaly profiles with real pathology profiles rather than random noise (ablation) | UNA · Simulated pathology, T1w MRI, full brain | L1: 0.0195 -> 0.0147 | Helped |
| add contralateral-paired input (ablation) | UNA · Simulated pathology, T1w MRI, full brain | L1: 0.0155 -> 0.0147 | Helped |
| add intra-subject self-contrastive loss (ablation) | UNA · Simulated pathology, T1w MRI, full brain | L1: 0.0195 -> 0.0147 | Helped |
Surprising findings
- Ablation without any fluid-driven randomization (training only on real pathology images) gives only slight gains over Brain-ID, which was never trained on diseased inputs.
- UNA tends to over-correct diseased anatomy when pathology heavily occludes the underlying structure.
Limitations
- Over-corrects healthy anatomy in extreme cases where pathology occludes underlying anatomy
- Relies on contralateral symmetry of brain anatomy and precomputed NiftyReg deformations
- Evaluated mainly on brain stroke data; broader applications such as registration and outcome prediction not yet explored
Experimental setup
Datasets: ADNI ADNI3 HCP ADHD200 AIBL OASIS3 ATLAS ISLES
Benchmarks: ADNI ADNI3 HCP ADHD200 AIBL OASIS3 ATLAS ISLES simulated pathology test set (1,000 samples)
Metrics: L1 PSNR SSIM Dice
Baselines: SynthSR Brain-ID PEPSI VAE LDM
Builds on: SynthSR Brain-ID PEPSI advection-diffusion PDE NiftyReg
Closest papers by meaning
- NOVA: A Benchmark for Rare Anomaly Localization and Clinical Reasoning in Brain MRI (NeurIPS 2025 · similarity 0.86)
- STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation (NeurIPS 2026 · similarity 0.85)
- Anatomical Knowledge Mining and Matching for Semi-supervised Medical Multi-structure Detection (AAAI 2025 · similarity 0.84)
- Leveraging Anatomical Consistency for Multi-Object Detection in Ultrasound Images via Source-free Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.84)
- Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images (ICML 2024 · similarity 0.84)
- Point2SSM: Learning Morphological Variations of Anatomies from Point Clouds (ICLR 2024 · similarity 0.83)
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.