Automatic Diagnosis of Spasmodic Dysphonia with Structural ...Spasmodic Dysphonia (SD) is a...
1
Spasmodic Dysphonia (SD) is a neurological disorder of unknown pathophysiology, characterized by involuntary spasms in laryngeal muscles predominantly during speaking. Neural markers of SD are non-existent, resulting in significant delays in diagnosis and treatment of this disorder. Background – Diagnosing spasmodic dysphonia is a challenging task Automatic Diagnosis of Spasmodic Dysphonia with Structural MRI and Machine Learning Davide Valeriani, Ph.D., and Kristina Simonyan, M.D., Ph.D., Dr. med. [email protected] – [email protected] Methods – Two alternative pipelines for selecting classification features Machine learning Diagnosis Structural MRI SD Healthy o Supervised pipeline Gray matter volume (GMV) Meta-analysis (GingerALE) on six SD studies for feature selection mean GMV and CT in each cluster Ensemble of convolutional neural networks (CNNs) $ Funding Aim – Identify objective markers of SD with structural MRI and machine learning • 104 high-resolution T1-weighted images (52 SD patients and 52 controls) • Two pipelines for the identification of diagnostic markers • Four machine-learning (ML) algorithms for diagnostic classification Cortical thickness (CT) T1-weighted image Semi-supervised pipeline Conv3D (5x5x5) Conv3D (3x3x3) Conv3D (3x3x3) Conv3D (3x3x3) Dense Dense Dense Dense MaxPooling Dropout 0.3 MaxPooling Dropout 0.3 Dropout 0.5 Dropout 0.5 Dropout 0.5 Conv3D (5x5x5) 64@25x25x25 Conv3D (3x3x3) 128@25x25x25 Conv3D (3x3x3) 128@9x9x9 Conv3D (3x3x3) 256@9x9x9 Dense 512 Dense 256 Dense 128 Dense 1 MaxPooling Dropout 0.3 MaxPooling Dropout 0.3 Dropout 0.5 Dropout 0.5 Dropout 0.5 + 13-fold cross-validation Conclusions and Future Work • High accuracy of objective diagnosis based on the use of both supervised and semi-supervised ML and neural alterations in spasmodic dysphonia • Accuracy of ML-based diagnosis outperforms more than two-fold the diagnostic agreement between physicians: 72% (this study) vs. 34% (Ludlow et al., 2018) • Future: Larger sample size may improve the accuracy of CNNs Input and preprocessing AUC = 72.2% AUC = 70.2% AUC = 66.4% AUC = 68.0% Support vector machines (SVM) linear kernel and C=100 Neural network (NN) one hidden layer, 12 neurons Linear discriminant analysis (LDA) Ensemble of CNNs Results – Objective markers of SD diagnosis #1 y=12 Primary motor cortex #2 y=-2 Premotor cortex #5 y=4 #6 y=10 Inferior frontal gyrus Primary motor cortex #3 y=10 Putamen #4 y=40 Inferior parietal cortex LDA SVM NN test set: 25% Volumetric patches Poster number: W277 Generous gift by Mr. and Mrs. Richardson
Transcript of Automatic Diagnosis of Spasmodic Dysphonia with Structural ...Spasmodic Dysphonia (SD) is a...
Spasmodic Dysphonia (SD) is a neurological disorder of unknown pathophysiology, characterized by involuntary spasms in laryngeal muscles predominantly during speaking. Neural markers of SD are non-existent, resulting in significant delays in diagnosis and treatment of this disorder.
Background – Diagnosing spasmodic dysphonia is a challenging task
Automatic Diagnosis of Spasmodic Dysphonia with Structural MRI and Machine Learning
Davide Valeriani, Ph.D., and Kristina Simonyan, M.D., Ph.D., Dr. [email protected] – [email protected]
Methods – Two alternative pipelines for selecting classification features
Machine learning DiagnosisStructural MRI
SD
Healthy
o
Supervised pipeline
Gray matter volume (GMV)
Meta-analysis (GingerALE) on six SD studies for feature selection
mean GMV and CT in each cluster
Ensemble of convolutional neural networks (CNNs)
$ Funding
Aim – Identify objective markers of SD with structural MRI and machine learning• 104 high-resolution T1-weighted images
(52 SD patients and 52 controls)• Two pipelines for the identification of
diagnostic markers• Four machine-learning (ML) algorithms
for diagnostic classification
Cortical thickness (CT)
T1-weighted image
Semi-supervised pipeline
Conv3D(5x5x5)
Conv3D(3x3x3)
Conv3D(3x3x3)
Conv3D(3x3x3) De
nse
Dens
e
Dens
e
Dens
e
Max
Pool
ing
Drop
out 0
.3
Max
Pool
ing
Drop
out 0
.3
Drop
out 0
.5
Drop
out 0
.5
Drop
out 0
.5
Conv3D(5x5x5)
64@25x25x25
Conv3D(3x3x3)
128@25x25x25
Conv3D(3x3x3)
128@9x9x9
Conv3D(3x3x3)
256@9x9x9
Dens
e
512
Dens
e
256
Dens
e
128
Dens
e
1
Max
Pool
ing
Drop
out 0
.3
Max
Pool
ing
Drop
out 0
.3
Drop
out 0
.5
Drop
out 0
.5
Drop
out 0
.5
+
13-fold cross-validation
Conclusions and Future Work• High accuracy of objective diagnosis based on the use of both supervised and
semi-supervised ML and neural alterations in spasmodic dysphonia• Accuracy of ML-based diagnosis outperforms more than two-fold the diagnostic
agreement between physicians: 72% (this study) vs. 34% (Ludlow et al., 2018)• Future: Larger sample size may improve the accuracy of CNNs
Input and preprocessing
AUC = 72.2% AUC = 70.2%
AUC = 66.4% AUC = 68.0%
Support vector machines (SVM) linear kernel and C=100
Neural network (NN) one hidden layer, 12 neurons
Linear discriminant analysis (LDA)
Ensemble of CNNs
Results – Objective markers of SD diagnosis
#1 y=12
Primary motor cortex
#2 y=-2
Premotor cortex#5 y=4 #6 y=10
Inferior frontal gyrus
Primary motor cortex
#3 y=10
Putamen#4 y=40
Inferior parietal cortex
LDA SVM NN
test set: 25%
Volumetric patches
Poster number: W277
Generous gift by Mr. and Mrs. Richardson
