CRIS: Cross-Plane Self-Supervised Isotropic Restoration for Anisotropic Volumetric Imaging Across Modalities

2026-06-14 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Imaging · Depth: Expert, quick

Summary

CRIS, a cross-plane self-supervised framework, addresses the challenge of anisotropic volumetric imaging in clinical MRI and volume electron microscopy (vEM) by restoring isotropic resolution without requiring paired isotropic ground truth. Anisotropic acquisitions, common in these modalities, produce thick slices that degrade orthogonal reformats. CRIS frames 3D restoration as 2D stripe completion on orthogonal reformats of an isotropic grid, synthetically degrading and masking high-resolution in-plane slices for training, then fusing predictions via multi-view averaging during inference. Evaluated on two MRI cohorts and two microscopy benchmarks up to 8x anisotropy, CRIS achieved 32.921 +/- 0.436 dB PSNR and 0.9631 +/- 0.0027 SSIM on brain MRI, outperforming several existing methods. It also reduced FID/KID to 48.714/0.023 on abdominal MRI and demonstrated superior performance on vEM benchmarks, reaching 29.133 dB/0.834 3D PSNR/SSIM at 4x. The framework is modality-flexible and avoids configuration-specific retraining.

Key takeaway

For Machine Learning Engineers developing medical imaging or microscopy pipelines, CRIS offers a robust solution for isotropic restoration of anisotropic volumetric data. You can achieve superior image quality and segmentation consistency, as demonstrated by 32.921 dB PSNR on brain MRI, without needing paired isotropic ground truth or extensive retraining. Consider integrating CRIS to enhance downstream analysis and improve diagnostic accuracy across diverse modalities.

Key insights

CRIS enables isotropic volumetric image restoration using a self-supervised, cross-plane approach without paired ground truth.

Principles

• Anisotropic restoration can be framed as 2D stripe completion.
• Multi-view averaging improves prediction fusion.
• Self-supervision eliminates need for paired isotropic data.

Method

CRIS casts 3D restoration as 2D stripe completion on orthogonal reformats. It synthetically degrades and masks in-plane slices for training, then restores two orthogonal reformats and fuses predictions by multi-view averaging.

In practice

• Apply CRIS to clinical MRI for improved image quality.
• Use CRIS for volume electron microscopy analysis.
• Integrate CRIS for robust anisotropic data processing.

Topics

• Volumetric Imaging
• Isotropic Restoration
• Self-Supervised Learning
• Medical MRI
• Electron Microscopy
• Image Reconstruction

Code references

• adi-hatav/CRIS

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Universal restoration of medical images — HorusEye is a self-supervised foundation model for universal medical image restoration without clean training data.
• MNet++: Extended 2D/3D Networks for Anisotropic Medical Image Segmentation — Adaptive 2D-3D feature fusion and state-space models enhance MNet's anisotropic medical image segmentation reliability and reproducibility.
• 3D Consistency Optimization for Self-Supervised Monocular Video Depth Estimation — Recasting monocular video depth as multi-view 3D reconstruction with 3D consistency optimization improves geometric accuracy and coherence.
• Diffusion Autoencoder for Unsupervised Artifact Restoration in Handheld Fundus Images — An unsupervised diffusion autoencoder restores fundus image artifacts by learning semantic representations from high-quality data.
• Robust Multispectral Semantic Segmentation under Missing or Full Modalities via Structured Latent Projection — Structured latent projection in CBC-SLP preserves modality-specific information, enhancing multimodal semantic segmentation robustness.
• Support-Conditioned Flow Matching Is Kernel Smoothing — Cross-attention conditioning in generative models is kernel smoothing, with predictable failure modes and learned corrections.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.