Multimodal MRI-based classification of migraine: using deep learning convolutional neural network

DOI: https://doi.org/10.1186/s12938- 018- 0587- 0

Abstract-Summary Deep learning technologies have rapidly expanded into medical image analysis, including both disease detection and classification.

No studies have evaluated the potential of deep learning technologies in assisting

with the classification of migraine patients.

We used deep learning methods in combination with three functional measures (the amplitude of low-frequency fluctuations, regional homogeneity and regional functional correlation strength) based on rs-fMRI data to distinguish not only between migraineurs and healthy controls, but also between the two subtypes of migraine.

We employed 21 migraine patients without aura, 15 migraineurs with aura, and

28 healthy controls.

Compared with the traditional support vector machine classifier, which has an accuracy of 83.67%, our Inception module-based convolutional neural network approach showed a significant improvement in classification output (over 86.18%).

568

3 Diagnosis

Our study shows that combining the three functional measures of rs-fMRI with deep learning classification is a powerful method to distinguish between migraineurs and healthy individuals.

Our data also highlight that deep learning-based frameworks could be used to develop more complicated models or systems to aid in clinical decision making in the future.

Extended: This is one of the first studies to examine the performance of different deep learning-based frameworks and to apply them specifically for migraine discrimination.

We used two deep learning-based classifiers to distinguish between healthy brains, brains affected by MWA and brains affected by MWoA. The first model we used was the CNN network based on AlexNet [472], and the second model was the CNN with Google’s Inception module.

Compared with the support vector machine classifier that we previously ana- lyzed (a final classification accuracy of 83.67%) [473], our approach provides pre- liminary support for deep learning methods combined with the fMRI features as a method for improving the discriminative power for migraine.

Background By using resting-state fMRI, researchers have demonstrated that migraines are related to different indices of functional brain alterations, including amplitude of low-frequency fluctuations (ALFF), regional homogeneity (ReHo), and regional functional correlation strength (RFCS).

Despite these results demonstrating that migraines might contribute to functional brain alterations due to the repetitive occurrence of pain-related processes, very few studies have considered the possibility of using these functional features to improve the classification and diagnosis of migraines.

We used two deep learning-based classifiers to distinguish between healthy brains, brains affected by MWA and brains affected by MWoA. The first model we used was the CNN network based on AlexNet [472], and the second model was the CNN with Google’s Inception module.

This is one of the first studies to examine the performance of different deep them specifically for migraine

to apply

learning-based frameworks and discrimination.

Methods Imaging data were collected transversely by using an echo-planar imaging (EPI) sequence with the following settings: repetition time/echo time (TR/TE) = 1900/2.26 ms, flip angle  =  9°, slice thickness/gap  =  1/0 mm, field of view (FOV)  =  256 × 256 mm2, matrix = 256 × 256, and voxel size = 1 × 1×1 mm3.

The rs-fMRI data were also collected using an echo planar imaging (EPI) sequence but with the following settings: TR/TE = 2000/30 ms, flip angle = 90°, slice thickness/gap = 5/0 mm, FOV = 240 × 240 mm2, matrix = 64 × 64, and voxel size = 3.75 × 3.75 × 5 mm3.

3.4 Neuroimaging

569

With respect to the input layer, the post-feature mapping fMRI images were used as input data, with the convolutional layer playing the most important role in the CNN architectures as it is the core building block of such networks.

Results The best performer out of three different features was the RFCS, and the highest identification rate achieved was 99.25% when using the Inception module-based CNN to distinguish between the HC and migraine groups.

According to the experimental data, the recognition rate of the AlexNet-based CNN was lower than that of the Inception module-based CNN, especially in terms of the ALFF feature.

The Inception module-based CNN improved the classification performance in

most cases.

It was relatively hard for either framework to distinguish between the MWoA and MWA groups, with the AlexNet-based CNN showing an identification rate of just 86.43%.

Compared with other features, the RFCS feature mapping improved the classifi- cation accuracy of both of the deep learning-based models, with a noticeable differ- ence between the two.

Discussion We examined the ability of deep learning-based frameworks to discriminate between MWoA, MWA, and HC using features extracted from rs-fMRI data.

Compared with the support vector machine classifier that we previously ana- lyzed (a final classification accuracy of 83.67%) [473], our approach provides pre- liminary support for deep learning methods combined with the fMRI features as a method for improving the discriminative power for migraine.

We obtained an accuracy as high as 99.25% when using the RFCS feature in

deep learning-based frameworks.

We conclude that the deep learning-based frameworks can help identify migraine

patients when using these fMRI features.

For all features tested, the Inception module-based CNN resulted in a higher

level of accuracy than the AlexNet-based CNN.

Of these deep learning-based frameworks, our future goal is to visualize the

brain regions that are most affected by migraines.

Conclusion RFCS, ReHo and ALFF, these three functional indices we employed, can be used to represent different degrees of classification features.

Acknowledgement A machine generated summary based on the work of Yang, Hao; Zhang, Junran; Liu, Qihong; Wang, Yi, 2018 in BioMedical Engineering OnLine

570

3 Diagnosis

3.5

Medication Overuse and Addiction

Machine generated keywords: moh, medication, overuse, medicationoveruse, medi- cationoveruse headache, moh patient, drug, medication overuse, patient moh, acute medication, substance, triptan, addiction, gray, withdrawal

Medication overuse and drug addiction: a narrative review from addiction perspective