[Paper] Adaptive Conditional Contrast-Agnostic Deformable Image Registration with Uncertainty Estimation

Source: arXiv - 2601.05981v1

Overview

The paper introduces Adaptive Conditional Contrast‑Agnostic Registration (AC‑CAR), a deep‑learning framework that can align medical images taken with any imaging contrast—without ever having seen that contrast during training. By combining a clever contrast‑augmentation trick with an adaptive feature‑modulation module, AC‑CAR not only registers images faster than classic iterative methods but also provides a built‑in estimate of registration uncertainty, making it far more trustworthy for downstream clinical or research pipelines.

Key Contributions

• Contrast‑agnostic registration: A random‑convolution based augmentation scheme lets the network learn features that are invariant to any MRI/CT contrast, enabling zero‑shot generalization to unseen modalities.
• Adaptive Conditional Feature Modulator (ACFM): Dynamically adjusts feature maps based on the inferred contrast, enforcing consistency across disparate image appearances.
• Latent regularization for invariance: Introduces a contrast‑invariant latent loss that penalizes feature drift when the same anatomy is presented under different contrasts.
• Uncertainty estimation: Integrates a variance network that leverages the same encoder to output voxel‑wise confidence maps, improving interpretability and safety.
• Open‑source implementation: Full code and pretrained models are released (GitHub), facilitating reproducibility and rapid adoption.

Methodology

1. Contrast Augmentation via Random Convolution – During training, each input image is passed through a randomly generated convolution filter that mimics the intensity transformation of different imaging contrasts. This forces the network to treat contrast changes as a form of data augmentation rather than a new domain.
2. Encoder‑Decoder Registration Backbone – A standard U‑Net style encoder extracts multi‑scale features, while a decoder predicts a dense deformation field that warps the moving image onto the fixed image.
3. Adaptive Conditional Feature Modulator (ACFM) – The encoder also outputs a compact “contrast code”. The ACFM uses this code to scale and shift intermediate feature maps, effectively normalizing away contrast‑specific information.
4. Contrast‑Invariant Latent Regularization – The latent representation of the same anatomical pair, seen under two different random convolutions, is forced to be close in L2 distance, encouraging the network to learn contrast‑agnostic embeddings.
5. Uncertainty (Variance) Network – Parallel to the deformation decoder, a lightweight head predicts per‑voxel variance. During inference, the variance map can be thresholded or visualized to flag low‑confidence regions.
6. Training Objective – The loss combines (i) a similarity term (e.g., NCC or MSE) on the warped image, (ii) a smoothness regularizer on the deformation field, (iii) the latent invariance loss, and (iv) a negative log‑likelihood term that ties the variance predictions to the registration error.

Results & Findings

• Speed: Inference runs in ~30 ms per 3D volume on a single RTX 3090, compared to 10–30 s for classic iterative demons or B‑spline methods.
• Uncertainty Calibration: Regions with high predicted variance correlate strongly (ρ ≈ 0.71) with actual registration error, allowing downstream pipelines to automatically down‑weight unreliable voxels.
• Ablation: Removing the ACFM drops Dice by ~4 % on unseen contrasts, confirming its role in contrast adaptation.

Practical Implications

• Plug‑and‑play registration service: Developers can integrate AC‑CAR as a micro‑service that accepts any 3‑D medical volume (MRI, CT, PET) and returns a deformation field plus a confidence map—no need to retrain for each new contrast.
• Accelerated clinical workflows: Real‑time registration enables on‑the‑fly image fusion for image‑guided interventions, reducing patient time in the scanner.
• Robust multi‑center studies: Because the model generalizes across scanner protocols, researchers can harmonize heterogeneous datasets without manual preprocessing.
• Safety‑critical AI: The built‑in uncertainty estimate satisfies emerging regulatory expectations (e.g., FDA’s “trustworthiness” guidelines) by flagging regions where the model may be unreliable.
• Open‑source extensibility: The GitHub repo includes Dockerfiles and ONNX export scripts, making it straightforward to deploy on cloud platforms (AWS SageMaker, GCP AI Platform) or edge devices with GPU acceleration.

Limitations & Future Work

• Contrast space coverage: While random convolutions simulate many intensity mappings, extreme pathological contrasts (e.g., heavily attenuated CT with metal artifacts) were not evaluated and may still challenge the model.
• Memory footprint: The 3‑D U‑Net backbone with ACFM consumes ~8 GB GPU memory for typical 256³ volumes, limiting deployment on low‑end hardware.
• Uncertainty calibration: The variance network is trained jointly with registration; more rigorous Bayesian formulations could yield better calibrated confidence scores.
• Extension to non‑rigid modalities: Future work could explore integrating AC‑CAR with diffeomorphic constraints or incorporating biomechanical priors for organ‑specific deformation.

Overall, AC‑CAR pushes the frontier of contrast‑agnostic medical image registration, offering developers a fast, generalizable, and trustworthy tool ready for real‑world deployment.

Authors

• Yinsong Wang
• Xinzhe Luo
• Siyi Du
• Chen Qin

Paper Information

• arXiv ID: 2601.05981v1
• Categories: cs.CV
• Published: January 9, 2026
• PDF: Download PDF