FlowLet: Conditional 3D brain MRI synthesis using wavelet flow matching.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Generative modeling of 3D brain MRI faces challenges balancing anatomical accuracy, data diversity, and computational speed.
• Existing diffusion models offer high visual quality but require extensive sampling steps.
• Latent-space compression methods can introduce artifacts and compromise fine anatomical details.

Purpose of the Study:

• To introduce FlowLet, a conditional generative framework for efficient and accurate 3D brain MRI generation.
• To enable multi-scale generation without latent compression and fast inference via deterministic sampling.
• To provide explicit control over age-related morphological variations using novel conditioning techniques.

Main Methods:

• FlowLet utilizes Flow Matching in an invertible 3D wavelet domain for multi-scale generation.
• Deterministic Ordinary Differential Equation (ODE) sampling enables rapid inference.
• Age conditioning is implemented via feature-wise modulation and spatially adaptive cross-attention.

Main Results:

• FlowLet demonstrates competitive or superior global fidelity compared to diffusion baselines with significantly fewer sampling steps (as few as 10).
• Region-based evaluations confirm improved local anatomical plausibility across 95 brain regions.
• Augmenting training data with FlowLet-generated samples consistently reduces brain age prediction error.

Conclusions:

• FlowLet effectively addresses the trade-offs in 3D brain MRI generation, offering improved efficiency, controllability, and anatomical fidelity.
• The framework achieves anatomically meaningful generation, outperforming existing methods in key metrics.
• The open-source release of FlowLet aims to foster reproducibility and further advancements in neuroimaging research.