ReLiF-3D
MICCAI 2026 Early Accept Top 9% of Submissions

ReLiF-3D: Prior-Guided Semi-Supervised 3D MRI Segmentation via Robust Bias-Consistent Paired Views

Kunal Jangid, Tanmay Basu, Vinod Kurmi

AI Research · VisDom Lab · IISER Bhopal

Paper Code
Medical AI Foundation Models Semi-Supervised Learning 3D MRI Segmentation
Problem Few labeled 3D volumes

Manual voxel-level MRI annotation is expensive and hard to scale.

Idea Stable foundation prior

Use a frozen SAM-Med3D prior instead of a drifting EMA teacher.

Robustness Bias-consistent paired views

Model scanner-dependent shifts using smooth orthogonal bias fields.

Overview

Foundation-guided learning for low-label 3D MRI segmentation.

Overview diagram of the ReLiF-3D training pipeline
Overview of ReLiF-3D. A frozen SAM-Med3D prior guides a lightweight trainable 3D U-Net. SOBF generates bias-consistent paired views, while CGCR and LAR stabilize learning from limited labeled volumes.
Abstract

Stable prior supervision without a drifting teacher.

ReLiF-3D is a foundation-guided semi-supervised framework for 3D MRI segmentation under limited annotation. The method trains a lightweight 3D U-Net using supervision from a frozen volumetric foundation model prior, avoiding unstable EMA teacher drift while retaining a stable spatial guide. Its Smooth Orthogonal Bias Field (SOBF) generator produces anatomically consistent scanner-shifted paired views, and confidence-gated co-regularization enforces consistency only on reliable agreements. Lesion-aware representation alignment further improves robustness in low-label settings.

Frozen prior SAM-Med3D provides stable volumetric guidance.
Bias-consistent views SOBF models scanner-dependent MRI intensity shifts.
Reliable consistency CGCR and LAR reduce noisy pseudo-supervision.
Method

Four components form the ReLiF-3D training recipe.

01

Frozen SAM-Med3D Prior

Provides stable anatomical guidance without EMA teacher updates.

02

SOBF Paired Views

Simulates realistic scanner-dependent MRI bias through smooth multiplicative and additive fields.

03

Confidence-Gated Co-Regularization

Applies consistency constraints only where paired predictions agree reliably.

04

Lesion-Aware Alignment

Aligns lesion-relevant features with the foundation prior while reducing noisy background gradients.

SOBF Visual

Smooth orthogonal fields simulate scanner-dependent MRI bias.

SOBF constructs paired MRI views using smooth orthogonal multiplicative and additive fields. The visualization below shows clean images, generated bias fields, SOBF images, difference maps, and ground-truth masks.

SOBF visualization with clean images, M field, A field, SOBF image, difference map, and ground truth
Quantitative

Quantitative performance under limited labels.

1 labeled volume
BraTS76.41 Dice LA76.84 Dice

ReLiF-3D remains stable in the hardest low-label setting on BraTS and LA.

2 labeled volumes
BraTS78.26 Dice LA79.14 Dice

Adding one more labeled scan gives consistent overlap gains while preserving boundary quality.

3 labeled volumes
BraTS76.33 Dice LA79.50 Dice

ReLiF-3D maintains strong LA performance while keeping BraTS overlap competitive.

5 labeled volumes
BraTS77.69 Dice LA80.22 Dice

The method continues improving as labels increase, without losing robustness to scanner shifts.

Key insight

Across extreme low-label regimes, ReLiF-3D improves overlap while reducing boundary errors. The strongest gains appear in the 1-label setting, where stable foundation guidance and SOBF paired views reduce the failure modes common in teacher-student SSL.

Qualitative

Qualitative comparisons and attention maps.

Segmentation comparison between ReLiF-3D, SemiSAM+, ground truth, and input MRI image
Segmentation comparison against SemiSAM+.
Grad-CAM visualization comparing ReLiF-3D and SemiSAM+ against ground truth
Grad-CAM attention visualization.
Citation

Cite ReLiF-3D.

@inproceedings{jangid2026relif3d,
  title     = {ReLiF-3D: Prior-Guided Semi-Supervised 3D MRI Segmentation
               via Robust Bias-Consistent Paired Views},
  author    = {Jangid, Kunal and Basu, Tanmay and Kurmi, Vinod},
  booktitle = {Proceedings of MICCAI},
  year      = {2026}
}

Contact

For queries and collaborations, contact kunal24@iiserb.ac.in. or jangidkunal1999@gmail.com