A machine learning‐based quality control tool for functional MRI images

Abstract

AbstractBackgroundChanges in functional MRI (fMRI) have been associated with Alzheimer’s disease and other neurological disorders. Quality control (QC) of images prior to their computational analysis is a key step in a clinical trial analysis workflow, and it is traditionally implemented visually by image analysts. However, the number of images acquired from large clinical studies makes neuroimaging QC a difficult task that diverts skilled resources and introduces inter‐rater variability. This problem is accentuated in fMRI, where image analysts need to visually inspect 4D sequences, and although image processing reports can be used as aid, the problem of having to inspect each report remains.MethodWe implemented an automated fMRI QC process using a stacking classifier (STC), random forest + support vector machine, with imaging quality metrics (IQM) as inputs. The IQMs were estimated with MRIQC. We also added a designed IQM that measures how centred the participant’s brain is within the image. As train/test dataset we used the ABIDE multisite database. For our experiments, only adult brains from the ABIDE database were used, leaving 14 sites available for a leave‐one‐site‐out (LOSO) cross‐validation experiment. All images included in the experiment were visually QC’d using validated protocols at IXICO.ResultThe LOSO validation procedure resulted in a mean pass‐vs‐fail accuracy of 92% (with 5.7% standard deviation, SD). Pass/fail sensitivity and specificity resulted in 97% (6.7% SD) and 69% (29% SD) respectively. We tested a fully trained STC model on an independent dataset of healthy ageing participants from OpenNeuro (ds002872), which resulted in 87% accuracy, 94% sensitivity and 50% specificity, which validates our modelling procedure.ConclusionIt is possible to design and deploy an automatic QC model for fMRI. Although sensitivity was high in all models from the LOSO experiments, specificity was on average low, highlighting the difficulty of detecting fails in a QC pipeline. However, the STC model can be deployed to detect clear passes that will not need visual QC, reliably reducing the number of cases that require visual inspection. Our future work will focus on testing our developed model on additional datasets (e.g. ADNI) to validate its performance before roll‐out to clinical trial applications.aging after age 90.