Open Workflows (Lightning talk): AD-DL framework: A reproducible evaluation of convolutional neural networks for classification of Alzheimer's disease made easier with Clinica #39
AD-DL framework: A reproducible evaluation of convolutional neural networks for classification of Alzheimer's disease made easier with Clinica
By Alexandre Routier, Aramis Lab, Paris Brain Institute, Inria, Sorbonne University, Inserm, CNRS, France
Theme: Open Workflows
Format: Lightning talk
Abstract
Numerous machine learning approaches have been proposed for automatic classification of Alzheimer's disease (AD) from brain imaging data. However, the classification performance is difficult to compare across studies due to variations in components such as participant selection, image preprocessing or validation procedure. Moreover, these studies are hardly reproducible because their frameworks are not publicly accessible and because implementation details are lacking. Lastly, some of these approaches may report a biased performance due to inadequate or unclear validation or model selection procedures.
We present how to address these limitations through the combined use of the Clinica software and the AD-DL framework to evaluate convolutional neural networks (CNN) for AD classification from anatomical MRI. The corresponding papers has been published in Medical Image Analysis (see here for the post-print).
Clinica comprises open-source tools to automatically convert neuroimaging databases (ADNI, AIBL, NIFD and OASIS) into the BIDS standard, a modular set of image preprocessing procedures based on community software (ANTs, FreeSurfer, FSL, MRtrix, PETPVC, SPM), and pipelines to prepare data preprocessed by Clinica to be used with PyTorch. AD-DL proposes different CNN classification architectures based on the type of input (2D slice-level, 3D patch-level, ROI-based and 3D subject-level) and rigorous evaluation procedures.
AD-DL framework: A reproducible evaluation of convolutional neural networks for classification of Alzheimer's disease made easier with Clinica
By Alexandre Routier, Aramis Lab, Paris Brain Institute, Inria, Sorbonne University, Inserm, CNRS, France
Abstract
Numerous machine learning approaches have been proposed for automatic classification of Alzheimer's disease (AD) from brain imaging data. However, the classification performance is difficult to compare across studies due to variations in components such as participant selection, image preprocessing or validation procedure. Moreover, these studies are hardly reproducible because their frameworks are not publicly accessible and because implementation details are lacking. Lastly, some of these approaches may report a biased performance due to inadequate or unclear validation or model selection procedures.
We present how to address these limitations through the combined use of the Clinica software and the AD-DL framework to evaluate convolutional neural networks (CNN) for AD classification from anatomical MRI. The corresponding papers has been published in Medical Image Analysis (see here for the post-print).
Clinica comprises open-source tools to automatically convert neuroimaging databases (ADNI, AIBL, NIFD and OASIS) into the BIDS standard, a modular set of image preprocessing procedures based on community software (ANTs, FreeSurfer, FSL, MRtrix, PETPVC, SPM), and pipelines to prepare data preprocessed by Clinica to be used with PyTorch. AD-DL proposes different CNN classification architectures based on the type of input (2D slice-level, 3D patch-level, ROI-based and 3D subject-level) and rigorous evaluation procedures.
Contributors Junhao Wen
*(@anbai106) Elina Thibeau-Sutre*(@14thibea ) Mauricio Diaz-Melo (@mdiazmel) Jorge Samper-Gonzalez (@jsampergonzalez) Alexandre Routier (@alexandreroutier) Simona Bottani (@SimonaBottani) Didier Dormont Stanley Durrleman Ninon Burgos (@nburgos) Olivier Colliot (@oliviercolliot)*denotes shared first contributorsUseful Links
https://github.com/aramis-lab/AD-DL http://www.clinica.run/
Tagging @alexandreroutier