Automatic Diagnosis of Spasmodic Dysphonia with Structural MRI and Machine Learning

Poster No:

W277 

Submission Type:

Abstract Submission 

Authors:

Davide Valeriani1, Kristina Simonyan1

Institutions:

1Harvard Medical School, Boston, MA

E-Poster

Introduction:

Spasmodic dysphonia (SD) is a neurologic disorder of unknown causes and pathophysiology. SD is characterized by involuntary spasms in the laryngeal muscles that are specific to speech production. Highly variable symptomatology and the absence of objective diagnostic criteria make the diagnosis of SD challenging, leading to misdiagnosis and delayed treatment. As neuroimaging studies described SD-specific brain abnormalities contributing to its pathophysiology [Battistella et al., 2016, Simonyan and Ludlow, 2011], the aim of this study was to use a series of machine-learning algorithms to identify automatic objective diagnostic markers of SD based on structural abnormalities.

Methods:

Whole-brain T1-weighted MRI images of 52 SD patients (53.9±9.5 years, 33 females) and 52 age/gender-matched healthy controls (52.5±10.0 years, 33 females) were acquired on a 3T scanner. In each subject, FreeSurfer was used to extract cortical thickness (CT) and SPM12 CAT toolbox was used to extract gray matter volume (GMV). The performance of four classifiers was examined: (1) linear discriminant analysis (LDA), (2) linear support vector machines (SVM) with regularization strength C=100, (3) neural network (NN) with one hidden layer of 12 neurons, logistic activation, and Adam optimizer, and (4) an ensemble of convolution neural networks (CNNs).
For LDA, SVM and NN feature selection, we performed a separate meta-analysis (GingerALE) of previous imaging studies that applied voxel-based morphometry and CT analyses in SD patients and healthy controls [Simonyan and Ludlow, 2011, Bianchi et al, 2017, Kirke et al, 2017, Kostic et al, 2016, Termsarasab et al, 2016, Waugh et al, 2016]. Meta-analysis found six clusters at the voxel-wise significance level of p=0.001 and minimum cluster volume of 200 mm3. Using these clusters as a mask, the average CT and GMV measures were extracted, resulting in a total of 12 structural features per subject. The performance of LDA, SVM and NN was computed using a 13-fold cross-validation.
For CNN, the whole-brain GMV and CT images were subsampled by a factor of 2 and split into training (78 subjects) and test (26 subjects) sets. Data were augmented by extracting volumetric patches of 25x25x25 from each image. We trained two CNNs (Fig. 1) with Keras, one with CT images and one with GMV images. The outputs of the two CNNs were then averaged across patches associated to the same subject and across CNNs to obtain the ensemble prediction.
Supporting Image: Fig1.jpg
   ·Fig. 1. Architecture of the CNN.
 

Results:

Meta-analysis of VBM and CT literature in SD patients vs. healthy controls found significant structural abnormalities mainly present in the left hemisphere, including premotor cortex (#2), putamen (#3), inferior parietal cortex (#4), and inferior frontal gyrus (#5) (Fig. 2). Clusters in the primary motor cortex (#1 and #6) were found bilaterally.
Based on these data, the combination of CT and GMV features resulted in average cross-validation AUC of 72.6% for LDA, 70.2% for SVM, 66.4% for NN, and 53.3% for CNN.
Supporting Image: Fig2.jpg
   ·Fig. 2Coronal slices in standard Talairach-Tournoux space with clusters of significant abnormalities between SD patients and healthy controls identified by meta-analysis.
 

Conclusions:

Machine-learning classifiers based on meta-analysis driven disorder-specific brain abnormalities correctly diagnosed SD in approximately two out of three patients, showing their strong translational potential in contrast to clinical evaluation of dystonic symptoms with a 34% agreement rate among expert physicians [Ludlow et al, 2018]. LDA outperformed other algorithms and also showed superior performance to LDA applied to resting-state fMRI data in SD patients [Battistella et al, 2016]. This suggests that a combination of supervised structural feature selection and LDA may be a promising avenue for the development of objective tools for SD diagnosis. CNN suffered from underfitting due to the small sample size. Taken together, structural brain abnormalities identified in SD could serve as imaging markers for objective diagnosis of this disorder. Future studies should focus on integrating features extracted from other imaging modalities (e.g., fMRI) in the classification pipeline.

Disorders of the Nervous System:

Parkinson's Disease and Movement Disorders 1

Imaging Methods:

Anatomical MRI

Language:

Speech Production

Modeling and Analysis Methods:

Classification and Predictive Modeling 2

Keywords:

Machine Learning
Movement Disorder
STRUCTURAL MRI
Other - Dystonia

1|2Indicates the priority used for review

My abstract is being submitted as a Software Demonstration.

No

Please indicate below if your study was a "resting state" or "task-activation” study.

Other

Healthy subjects only or patients (note that patient studies may also involve healthy subjects):

Patients

Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.

Yes

Was any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.

Yes

Was any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.

Not applicable

Please indicate which methods were used in your research:

Structural MRI

For human MRI, what field strength scanner do you use?

3.0T

Which processing packages did you use for your study?

AFNI
SPM
Free Surfer
Other, Please list  -   FingerALE

Provide references using author date format

Battistella, G., Fuertinger, S., Fleysher, L., Ozelius, L. J., & Simonyan, K. (2016). Cortical sensorimotor alterations classify clinical phenotype and putative genotype of spasmodic dysphonia. European Journal of Neurology, 23(10), 1517-1527.
Simonyan, K., & Ludlow, C. L. (2011). Abnormal structure–function relationship in spasmodic dysphonia. Cerebral Cortex, 22(2), 417-425.
Bianchi, S., Battistella, G., Huddleston, H., Scharf, R., Fleysher, L., Rumbach, A. F., ... & Simonyan, K. (2017). Phenotype‐and genotype‐specific structural alterations in spasmodic dysphonia. Movement Disorders, 32(4), 560-568.
Kirke, D. N., Battistella, G., Kumar, V., Rubien-Thomas, E., Choy, M., Rumbach, A., & Simonyan, K. (2017). Neural correlates of dystonic tremor: a multimodal study of voice tremor in spasmodic dysphonia. Brain imaging and behavior, 11(1), 166-175.
Kostic, V. S., Agosta, F., Sarro, L., Tomić, A., Kresojević, N., Galantucci, S., ... & Filippi, M. (2016). Brain structural changes in spasmodic dysphonia: a multimodal magnetic resonance imaging study. Parkinsonism & related disorders, 25, 78-84.
Termsarasab, P., Ramdhani, R. A., Battistella, G., Rubien-Thomas, E., Choy, M., Farwell, I. M., ... & Hutchinson, M. (2016). Neural correlates of abnormal sensory discrimination in laryngeal dystonia. Neuroimage: Clinical, 10, 18-26.
Waugh, J. L., Kuster, J. K., Levenstein, J. M., Makris, N., Multhaupt-Buell, T. J., Sudarsky, L. R., ... & Blood, A. J. (2016). Thalamic volume is reduced in cervical and laryngeal dystonias. PloS one, 11(5), e0155302.
Ludlow, C. L., Domangue, R., Sharma, D., Jinnah, H. A., Perlmutter, J. S., Berke, G., ... & Blindauer, K. (2018). Consensus-based attributes for identifying patients with spasmodic dysphonia and other voice disorders. JAMA Otolaryngology–Head & Neck Surgery.