Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FlowLet: Conditional 3D Brain MRI Synthesis using Wavelet Flow Matching

Danilo Danese, Angela Lombardi, Matteo Attimonelli, Giuseppe Fasano, Tommaso Di Noia

TL;DR

FlowLet is proposed, a conditional generative framework that synthesizes age-conditioned 3D MRIs by leveraging flow matching within an invertible 3D wavelet domain, helping to avoid reconstruction artifacts and reducing computational demands.

Abstract

Brain Magnetic Resonance Imaging (MRI) plays a central role in studying neurological development, aging, and diseases. One key application is Brain Age Prediction (BAP), which estimates an individual's biological brain age from MRI data. Effective BAP models require large, diverse, and age-balanced datasets, whereas existing 3D MRI datasets are demographically skewed, limiting fairness and generalizability. Acquiring new data is costly and ethically constrained, motivating generative data augmentation. Current generative methods are often based on latent diffusion models, which operate in learned low dimensional latent spaces to address the memory demands of volumetric MRI data. However, these methods are typically slow at inference, may introduce artifacts due to latent compression, and are rarely conditioned on age, thereby affecting the BAP performance. In this work, we propose FlowLet, a conditional generative framework that synthesizes age-conditioned 3D MRIs by leveraging flow matching within an invertible 3D wavelet domain, helping to avoid reconstruction artifacts and reducing computational demands. Experiments show that FlowLet generates high-fidelity volumes with few sampling steps. Training BAP models with data generated by FlowLet improves performance for underrepresented age groups, and region-based analysis confirms preservation of anatomical structures.

FlowLet: Conditional 3D Brain MRI Synthesis using Wavelet Flow Matching

TL;DR

FlowLet is proposed, a conditional generative framework that synthesizes age-conditioned 3D MRIs by leveraging flow matching within an invertible 3D wavelet domain, helping to avoid reconstruction artifacts and reducing computational demands.

Abstract

Brain Magnetic Resonance Imaging (MRI) plays a central role in studying neurological development, aging, and diseases. One key application is Brain Age Prediction (BAP), which estimates an individual's biological brain age from MRI data. Effective BAP models require large, diverse, and age-balanced datasets, whereas existing 3D MRI datasets are demographically skewed, limiting fairness and generalizability. Acquiring new data is costly and ethically constrained, motivating generative data augmentation. Current generative methods are often based on latent diffusion models, which operate in learned low dimensional latent spaces to address the memory demands of volumetric MRI data. However, these methods are typically slow at inference, may introduce artifacts due to latent compression, and are rarely conditioned on age, thereby affecting the BAP performance. In this work, we propose FlowLet, a conditional generative framework that synthesizes age-conditioned 3D MRIs by leveraging flow matching within an invertible 3D wavelet domain, helping to avoid reconstruction artifacts and reducing computational demands. Experiments show that FlowLet generates high-fidelity volumes with few sampling steps. Training BAP models with data generated by FlowLet improves performance for underrepresented age groups, and region-based analysis confirms preservation of anatomical structures.
Paper Structure (76 sections, 17 equations, 8 figures, 11 tables)

This paper contains 76 sections, 17 equations, 8 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Augmentation
  4. Generative Models for 3D Synthesis
  5. Wavelet Diffusion Models (WDMs)
  6. Proposed Method
  7. Preliminaries
  8. FlowLet Framework
  9. Flow Matching Formulations
  10. Rectified Flow Matching (RFM)
  11. Conditional Flow Matching (CFM)
  12. Variance-Preserving Diffusion Matching (VP)
  13. Trigonometric Flow
  14. Conditional U-Net Architecture
  15. Experimental Setup
  16. ...and 61 more sections

Figures (8)

  • Figure 1: (a) Training in the wavelet domain decomposes the MRI into one low-frequency (LLL, black background) and seven high-frequency subbands. A conditional U-Net learns to predict velocity fields between noise and data. (b) Inference uses ODE integration followed by IDWT.
  • Figure 3: Visual assessment of image fidelity and realism for different 3D brain MRI synthesis models. Each column displays standard Axial, Coronal, and Sagittal views for the real reference scan (Subject of 72 years old) and specified generative method (Ours is RFM 10 steps).
  • Figure 4: Combined age distribution of the training dataset after integration of OpenBHB, ADNI, and OASIS-3.
  • Figure 5: Qualitative comparison of FlowLet flow matching formulation across different ODE step counts. Each column shows axial, coronal, and sagittal views for a given method and step count. All samples share the same noise seed and age condition, yielding anatomically consistent outputs with similar structural compartments across flow formulations.
  • Figure 6: FID vs. number of steps for FlowLet variants calculated on overall (5.9-95 years) and by age range. The shaded bands indicate standard deviation.
  • ...and 3 more figures