Classifiers trained on surface-based MRI features have been proposed to automatically detect subtle structural malformations, called focal cortical dysplasias (FCDs), which are amenable to surgery. To decrease chances of overfitting, these classifiers should be trained on large multi-centre datasets. However, systematic scanner differences can introduce site-specific biases in the data. Furthermore, heterogeneous patient cohorts mean that features can be affected by developmental effects such as age, as well as the biological variable of interest. To overcome these limitations, careful pre-processing of the data is needed. In this abstract, we present the pipeline used to post-process our Multi-centre Epilepsy Lesion Detection (MELD) cohort for the detection of FCDs.
The MELD cohort is composed of 555 patients and 390 controls, from 21 epilepsy centres worldwide. We extracted structural and intensity features from MRI images, e.g. cortical thickness and FLAIR intensity. Features were smoothed, harmonized across sites using ComBat (Fortin et al., 2018), and normalized for intersubject and interregional morphological differences.
This post-processing reduced site, scanner, age and cortical region -based biases in our data, enabling the quantification of subtle MRI feature differences related to FCD histopathologies. The MRI features were used to train a classifier to differentiate FCD subtypes with 80% accuracy.
We present a framework to harmonise large multi-centre MRI datasets. Moreover, this pipeline offers the flexibility to apply the normalisation on new data, enabling lesion detection on new patients and sites, which is integral for clinical utility in the presurgical evaluation of patients with epilepsy.