# Impact of AIR™ Recon DL on magnetic resonance imaging-based quantitative brain structure measurements

## TL;DR

This study compares a new MRI reconstruction algorithm with traditional methods and finds that while it improves image quality, it also causes small but consistent changes in brain volume measurements.

## Contribution

The study reveals systematic morphometric biases introduced by AIR™ Recon DL, a deep learning-based MRI reconstruction algorithm.

## Key findings

- AIR™ Recon DL significantly improves image quality by reducing noise and artifacts.
- The algorithm introduces systematic shifts in brain volume measurements, including increased gray matter and cerebrospinal fluid volumes.
- Reconstruction time is longer for AIR™ Recon DL compared to conventional reconstruction.

## Abstract

We aimed to evaluate how the AIR™ Recon DL algorithm influences magentic resonance imaging (MRI) quality and quantitative brain morphometry relative to conventional reconstruction (CR). Seventy-four healthy adults underwent 3D T1-weighted MRI reconstructed with CR and AIR™ Recon DL. Image quality was rated by two neuroradiologists (κ = 0.74–0.97). Voxel-based morphometry assessed total, gray matter (GM), white matter (WM), and cerebrospinal (CSF) volumes; surface-based morphometry analyzed cortical thickness, sulcal depth, fractal dimension, and gyrification across 148 regions. Hippocampal volumes were extracted using the Neuromorphometrics atlas. Reconstruction times were compared. AIR™ Recon DL significantly improved image quality (reduced noise and artifacts, P < 0.001) but introduced systematic morphometric shifts—smaller total and WM volumes, larger GM and CSF volumes, and widespread regional thickness increases (effect sizes d ≈ 0.3–0.5). Hippocampal volumes increased bilaterally (ΔL = +0.15 mL, +3.97%; ΔR = +0.15 mL, +3.88%; both P < 0.05). Mean reconstruction time was longer for deep learning-based reconstruction (11.6 ± 1.6 s) than CR (9.9 ± 1.4 s; Δ = +1.7 s, P < 0.001). AIR™ Recon DL enhances image quality but causes modest, systematic volumetric biases. Harmonizing reconstruction methods is essential for reliable morphometric comparisons in neuropsychiatric imaging.