# Easy-to-use and easy-to-interpret quality control of 3D gradient echo T1-weighted MR acquisition sequences for improved test-retest stability of MRI-based hippocampus volumetry

## TL;DR

This paper introduces a method to assess MRI scan quality for reliable hippocampus volume measurements in Alzheimer's research.

## Contribution

A novel decision tree model using image quality metrics to predict hippocampus volumetry stability in T1-weighted MRI scans.

## Key findings

- A CART model achieved 79.5% accuracy in classifying MRI sequences by hippocampus volumetry stability.
- Left-right FOV width and contrast-to-noise ratio were key predictors of scan quality.
- The model improved good-to-poor sequence ratio from 3.5 to 7.4 while losing 15% of good sequences.

## Abstract

MRI-based hippocampus volume (HV) is widely used as neurodegeneration marker in Alzheimer's disease.

An easy-to-use and easy-to-interpret method to categorize T1-weighted MR sequences with respect to test-retest stability of hippocampus volumetry based on general image quality metrics (IQM).

The study included 446 3D T1-weighted MRI scans of one healthy middle-aged man obtained during 32 months in 122 scanning sessions performed with 96 different scanners at 76 different sites. Each scanning session represented a different acquisition sequence of ≥2 back-to-back repeat scans (3.7 ± 0.7 on average). Unilateral HVs were determined with 18 different tools for automatic volumetry. An acquisition sequence was considered “poor” if the z-score of the within-session coefficient-of-variation of the HV estimates from the session, averaged across all volumetry tools and both hemispheres, exceeded one standard deviation. General IQM were computed for each scanning session using the freely available MRI Quality Control Tool. A classification-and-regression tree (CART) was trained to discriminate between good and poor acquisition sequences using the IQM as input.

The CART selected the left-right width of the acquisition field-of-view and the contrast-to-noise ratio as predictor variables. Overall accuracy of the CART was 79.5%. CART-based classification increased the ratio of good-to-poor acquisition sequences from 3.5 among all sequences to 7.4 among the sequences predicted to be good. This was at the expense of losing 15% of the good sequences.

The IQM-based decision tree model provides useful performance for the differentiation of T1-weighted sequences associated with good versus poor test-retest stability of hippocampus volumetry.