Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-modal MRI Translation via Evidential Regression and Distribution Calibration

Jiyao Liu, Shangqi Gao, Yuxin Li, Lihao Liu, Xin Gao, Zhaohu Xing, Junzhi Ning, Yanzhou Su, Xiao-Yong Zhang, Junjun He, Ningsheng Xu, Xiahai Zhuang

TL;DR

This work tackles the problem of translating missing MRI modalities across multi-modal scans while providing reliable uncertainty estimates and cross-center robustness. It reframes translation as an evidential regression problem and uses a Mixture of Normal-Inverse Gamma ($\mathrm{MoNIG}$) for explicit, uncertainty-aware fusion of source modalities, producing a joint predictive distribution. To cope with center-specific distribution shifts, a distribution calibration module based on quantile regression learns a center-adaptive mapping from the predictive CDF to calibrated distributions using a small calibration set. Experiments on BraTS2023 datasets (BraSyn, BraTS-Africa, BraTS-PED) demonstrate superior synthesis performance, calibrated uncertainty, and improved cross-domain robustness, with notable gains in downstream tumor segmentation metrics.

Abstract

Multi-modal Magnetic Resonance Imaging (MRI) translation leverages information from source MRI sequences to generate target modalities, enabling comprehensive diagnosis while overcoming the limitations of acquiring all sequences. While existing deep-learning-based multi-modal MRI translation methods have shown promising potential, they still face two key challenges: 1) lack of reliable uncertainty quantification for synthesized images, and 2) limited robustness when deployed across different medical centers. To address these challenges, we propose a novel framework that reformulates multi-modal MRI translation as a multi-modal evidential regression problem with distribution calibration. Our approach incorporates two key components: 1) an evidential regression module that estimates uncertainties from different source modalities and an explicit distribution mixture strategy for transparent multi-modal fusion, and 2) a distribution calibration mechanism that adapts to source-target mapping shifts to ensure consistent performance across different medical centers. Extensive experiments on three datasets from the BraTS2023 challenge demonstrate that our framework achieves superior performance and robustness across domains.

Multi-modal MRI Translation via Evidential Regression and Distribution Calibration

TL;DR

This work tackles the problem of translating missing MRI modalities across multi-modal scans while providing reliable uncertainty estimates and cross-center robustness. It reframes translation as an evidential regression problem and uses a Mixture of Normal-Inverse Gamma () for explicit, uncertainty-aware fusion of source modalities, producing a joint predictive distribution. To cope with center-specific distribution shifts, a distribution calibration module based on quantile regression learns a center-adaptive mapping from the predictive CDF to calibrated distributions using a small calibration set. Experiments on BraTS2023 datasets (BraSyn, BraTS-Africa, BraTS-PED) demonstrate superior synthesis performance, calibrated uncertainty, and improved cross-domain robustness, with notable gains in downstream tumor segmentation metrics.

Abstract

Multi-modal Magnetic Resonance Imaging (MRI) translation leverages information from source MRI sequences to generate target modalities, enabling comprehensive diagnosis while overcoming the limitations of acquiring all sequences. While existing deep-learning-based multi-modal MRI translation methods have shown promising potential, they still face two key challenges: 1) lack of reliable uncertainty quantification for synthesized images, and 2) limited robustness when deployed across different medical centers. To address these challenges, we propose a novel framework that reformulates multi-modal MRI translation as a multi-modal evidential regression problem with distribution calibration. Our approach incorporates two key components: 1) an evidential regression module that estimates uncertainties from different source modalities and an explicit distribution mixture strategy for transparent multi-modal fusion, and 2) a distribution calibration mechanism that adapts to source-target mapping shifts to ensure consistent performance across different medical centers. Extensive experiments on three datasets from the BraTS2023 challenge demonstrate that our framework achieves superior performance and robustness across domains.
Paper Structure (17 sections, 9 equations, 4 figures, 1 table)

This paper contains 17 sections, 9 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Uncertainty Estimation and Multi-modal Fusion
  4. Evidential Regression Module
  5. Uncertainty-guided Multi-modal Fusion
  6. Distribution Calibration
  7. Training Paradigm
  8. Experiments
  9. Dataset, Evaluation Metrics, and Implementation
  10. Datasets
  11. Evaluation Metrics
  12. Implementation Details
  13. Results
  14. Intra-domain Comparison Study
  15. Cross-domain Comparison Study
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Two challenges in multi-modal MRI translation: (1) lack of reliability assessment for synthesized images, and (2) limited robustness across domains.
  • Figure 2: Overview of the proposed framework. Through MoNIG-based image translation, we synthesize a target sequence from existing multi-modal source modalities based on explicit distribution fusion. Then, to adapt to the new source-target mapping, distribution calibration is performed to compute the calibrated generated target images and prediction uncertainty.
  • Figure 3: Visual comparisons of ours and other methods.
  • Figure 4: Left: Reliability visualization assessment, where purple mask shows enhanced tumor region, green mask shows oedema region, and red mask shows necrotic core. Right: Ablation study results evaluated on BraTS-Africa dataset.