Premium
Data‐assisted differential diagnosis of dementia by deep neural networks using MRI: A study from the European DLB consortium
Author(s) -
Larsen Simen Norrheim,
Oppedal Ketil,
Eftestøl Trygve,
Ferreira Daniel,
Lemstra Afina W.,
Kate Mara ten,
Padovani Alessandro,
Rektorova Irena,
Bonanni Laura,
Nobili Flavio Mariano,
Kramberger Milica G.,
Taylor JohnPaul,
Hort Jakub,
Snædal Jón,
Blanc Frédéric,
Antonini Angelo,
Borda Miguel German,
Aarsland Dag,
Westman Eric
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.043593
Subject(s) - differential diagnosis , dementia with lewy bodies , dementia , convolutional neural network , medical diagnosis , deep learning , artificial intelligence , neuroimaging , magnetic resonance imaging , medicine , computer science , disease , pathology , radiology , psychiatry
Background Clinicopathological overlap between neurodegenerative diseases contributes to challenging differentiation between dementia types. Improving differential diagnosis is important to provide optimal patient care and better predict future needs. Deep learning techniques such as convolutional neural networks have in recent years gained much attention for their high performance in image classification. Magnetic resonance imaging (MRI) computer‐assisted diagnosis (CAD) systems based on deep learning could potentially contribute to increased reliability of clinical diagnosis. Method An ensemble model of 3D‐convolutional neural networks was trained and tested on a class‐balanced dataset consisting of T1‐weighted MRI brain scans for the differentiation of subjects diagnosed with Alzheimer’s disease (AD), dementia with Lewy bodies (DLB) and normal controls (NC). Total number of brain scans used for training and testing were 690 (nAD=nDLB=nNC=230) and 171 (nAD=nDLB=nNC=57) respectively. Diagnoses were clinical, not including MRI, and in a subgroup of the DLB patients the diagnosis was supported by DAT scan. All scans were skull‐stripped and spatially normalized. All classes were matched for age and sex. Data augmentation was applied using the left and right hemisphere as separate instances and a Bayesian approach was used for hyperparameter optimization. Experiments were conducted using 6‐fold cross‐validation, with the final model constructed as an ensemble of all models. Result A final accuracy of 71.9% was reported for the three‐class differential classification of 171 stand‐out test subjects. The model classified the NC, DLB and AD subjects with 71.9%, 86.0% and 57.9% accuracy respectively. Further inspection showed that most of the miss‐classified AD patients were wrongly classified as DLB by the CAD system. Conclusion The results of this work show promising potential for deep learning based CAD systems trained on MR brain images for differential diagnosis of dementia. Most data‐driven dementia classification models, including the one presented in this work, are limited by the amount of available data, comorbidity, and the inaccuracy of clinical diagnosis. Increasing the amount of data might significantly improve such models. The prospective E‐DLB study will provide more brain MRI data for future improvement of the suggested deep learning based CAD system.