2689

CNN-based autoencoder and machine learning model for identifying betel-quid chewers using functional MRI features
Hsin-An Shen1, Ming-Chou Ho2,3, and Jun-Cheng Weng1,4,5
1Department of Medical Imaging and Radiological Sciences, and Bachelor Program in Artificial Intelligence, Chang Gung University, Taoyuan, Taiwan, 2Department of Psychology, Chung Shan Medical University, Taichung, Taiwan, 3Clinical Psychological Room, Chung Shan Medical University Hospital, Taichung, Taiwan, 4Medical Imaging Research Center, Institute for Radiological Research, Chang Gung University and Chang Gung Memorial Hospital at Linkou, Taoyuan, Taiwan, 5Department of Psychiatry, Chang Gung Memorial Hospital, Chiayi, Taiwan

Synopsis

Previous studies indicated that betel-quid chewing may cause brain functional alternations, but it cannot be distinguished with human eyes. We used resting-state functional magnetic resonance imaging as input features for machine learning to classify betel-quid chewers, alcohol- and tobacco-user controls, and healthy control.The results showed that logistic regression has a significant performance on identifying betel-quid chewers. The major advantage to this study is providing a 3D-autoencoder model and machine learning algorithm that can be used to discover the brain alternations in betel-quid chewers for clinical use in the future.

Introduction

Betel-quid (BQ) is one of the most commonly used psychoactive substances worldwide with especially high usage in Asian countries1. Although growing evidence indicates brain functional alterations and structural brain abnormalities in BQ chewers2-3, it is essentially impossible for radiologist to differentiate the MR images of BQ chewers from others with their eyes. Novel analytical methods, such as machine learning, can be used to develop algorithms to identify many diseases4. In this study, we used convolutional neural network (CNN)-based autoencoder model and machine learning algorithms to distinguish the BQ chewers base on the resting-state functional magnetic resonance imaging (rs-fMRI) features.

Methods

16 BQ chewers, 15 tobacco- and alcohol-user controls (TA), and 17 healthy controls (HC) were scanned using a 3-T MRI (Skyra, Siemens, Germany) imaging system with an echo-planar image (EPI) sequence to obtain resting-state functional MR images. The scanning protocol was performed with the following parameters: TR/TE = 2000/30 ms, field of view (FOV) = 250 mm× 250 mm2, matrix size = 94 × 94, in-plane resolution (pixel size) = 2.7 × 2.7 mm2, thickness = 4 mm, number of repetitions = 240, and 28 axial slices aligned along AC-PC lines without gap to cover the whole cerebrum. To increase the amount of data, we used different phase-encoding directions which are along the right-left (RL) and anterior-posterior (AP). The final total fMRI images included 33 HC, 30 TA, and 30 BQ chewers.
For fMRI preprocessing, statistical parametric mapping (SPM, Wellcome Department of Cognitive Neurology, London, UK) software was used. The mean fraction Amplitude of low-frequency fluctuations (mfALFF)5, 6 and the mean Regional homogeneity (mReHo)7 were then calculated using the Resting-State Data Analysis tool kit v1.8 (REST v1.8, Center for Cognition and Brain Disorders, Hangzhou Normal University, Zhejiang, China). In this analysis, we adopted a 3D autoencoder for feature selection in the fMRI datasets which included mfALFF, mReHo maps of the HC, TA, and BQ chewers (33, 30, 30 maps, respectively). Following the feature selection, the resulting compressed images were used for multiclass classification. We adopted 9 machine learning models, including (1) logistic regression (LR); (2) XGBoost (XGB); (3) decision tree classifier (CART); (4) linear discriminant analysis (LDA); (5) Gaussian naive Bayes (NB); (6) k-nearest neighbors classifier (KNN); (7) support vector machine (SVM); (8) multilayer perceptron (MLP); and (9) random forest (RF), and leave-one-out-cross-validation (LOOCV) was performed to find the best classifier for this research.

Results

LR reached the highest accuracy, which is 75% with mfALFF and 83% with mReHo. The results showed that LR had an impressive performance on classifying healthy controls, tobacco- and alcohol-user controls, and BQ chewers mutually exclusive using rs-fMRI as input features. Besides the accuracy, the correct classification rate of each category and Cohen’s kappa coefficient (Kappa)8 were also shown in Table 1.

Discussion

Although we constructed an autoencoder and supervised machine learning model for identifying BQ chewers, it suffers from the same limitations associated with high dimensionality of the feature sets and shortage of data. After the 3D-CNN compression for feature extraction, it only reduced the dimensionality but did not get better performance. The ratio of our training samples to dimensionality is low and it may lead to overfitting9. To prevent this, we applied LOOCV, which is a suitable method for a small dataset. Further work is required to address these complexities in classification.

Conclusion

The results from the present study showed that the machine learning algorithm, LR, was able to identify BQ chewers from tobacco- and alcohol-user controls, and healthy controls based on the data from rs-fMRI that cannot be differentiated by the human eye directly. It might provide a helpful approach to tracking BQ chewers or other brain alteration situation in the future.

Acknowledgements

This study was supported by the research programs, MOST109-2410-H-040-005 and MOST108-2410-H-040-005, which were sponsored by the Ministry of Science and Technology, Taipei, Taiwan.

References

1. Winstock AR (2002) Areca nut—abuse liability, dependence and public health. Addict Biol 7(1):133–138.

2. Upadhyay J, Maleki N, Potter J, Elman I, Rudrauf D, Knudsen J, WallinD, Pendse G, McDonald L, Griffin M, Anderson J, Nutile L,Renshaw P, Weiss R, Becerra L, Borsook D (2010) Alterations inbrain structure and functional connectivity in prescription opioiddependentpatients. Brain 133(7):2098–2114.

3. Chen F, Zhong Y, Zhang Z, Xu Q, Liu T, Pan M, Li J, Lu G (2015) Graymatter abnormalities associated with betel quid dependence: a voxel-based. morphometry study. Am J Transl Res 7(2):364–374.

4. Lundervold AS, Lundervold A. "An overview of deep learning in medical imaging focusing on MRI. " Zeitschrift fur Medizinische Physik. 2019 May;29(2):102-127.

5. Yue Y, Jia X, Hou Z, Zang Y, Yuan Y (2015) Frequency-dependentamplitude alterations of resting-state spontaneous fluctuations inlate-onset depression. Biomed Res Int 2015:505479.

6. Zou QH, Zhu CZ, Yang Y, Zuo XN, Long XY, Cao QJ, Wang YF, ZangYF (2008) An improved approach to detection of amplitude of lowfrequencyfluctuation (ALFF) for resting-state fMRI: fractional ALFF. J Neurosci Methods 172(1):137–141.

7. Zang Y, Jiang T, Lu Y, He Y, Tian L (2004) Regional homogeneityapproach to fMRI data analysis. NeuroImage 22(1):394–400.

8. Ben-David, Arie. (2008). Comparison of classification accuracy using Cohen’s Weighted Kappa. Expert Systems with Applications. 34. 825-832. 10.1016/j.eswa.2006.10.022.

9. Ying, Xue. (2019). An Overview of Overfitting and its Solutions. Journal of Physics: Conference Series. 1168. 022022.

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
2689