Single-Landmark vs. Multi-Landmark Deep Learning Approaches to Brain MRI Landmarking: a Case Study with Healthy Controls and Down Syndrome Individuals
Jordi Malé (La Salle - Ramon Llull University),* Yann Heuzé (CNRS, Univ. Bordeaux, MC, PACEA, UMR5199), Juan Fortea (Hospital of Sant Pau), Neus Martinez Abadias (Universitat de Barcelona), Xavier Sevillano (La Salle - Universitat Ramon Llull)
The 34th British Machine Vision Conference
Abstract
Brain dysmorphologies are present in many neurodegenerative, neurodevelopmental and genetic disorders, such as Alzheimer’s Disease, schizophrenia, or Down syndrome. Magnetic resonance imaging (MRI) is a widely used tool for diagnosing and monitor- ing these conditions, but the interpretation of MRI data can be challenging and time- consuming, particularly for large datasets. For this reason, there is a growing interest in developing automatic methods to detect anatomical landmarks in brain MRI data to quantify such dysmorphologies and obtain brain biomarkers to assist in the diagnosis and prognosis of these disorders. In this paper, we propose and evaluate two brain landmark- ing architectures based on Deep Convolutional Neural Networks (DCNN): an ensemble of single-landmark models, and a multi-landmark model. Both approaches are compared on MRI scans of healthy and Down syndrome subjects. The proposed pipeline comprises several steps: i) the preprocessing of the MRI data, involving registration to a common anatomical space, ii) the automatic extraction of the Mid-Sagittal Plane (MSP) of the brain, based on a multi-scale search algorithm, and iii) the training and evaluation of the DCNN models to detect 8 anatomical landmarks on the MSP. Our results indicate that i) the ensemble of single-landmark models is more accurate, achieving average landmark- ing errors lower than 2mm in healthy subjects, and ii) landmarking error is higher in Down syndrome individuals, which suggests that brain dysmorphologies associated with certain disorders require training specific models for accurate landmarking.Citation
@inproceedings{Malé_2023_BMVC,
author = {Jordi Malé and Yann Heuzé and Juan Fortea and Neus Martinez Abadias and Xavier Sevillano},
title = {Single-Landmark vs. Multi-Landmark Deep Learning Approaches to Brain MRI Landmarking: a Case Study with Healthy Controls and Down Syndrome Individuals},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://bmvc2022.mpi-inf.mpg.de/BMVC2023/0754.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by