کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Contrastive Learning Framework for fMRI Time-Series Classification in Left and Right Epilepsy Using Continues Wavelet Transform

نویسندگان :

Marzieh Soheili-nejad¹ Saeed Masoudnia² Hamid Soltanian-zadeh³

1- دانشگاه تهران 2- دانشگاه تهران 3- دانشگاه تهران

کلمات کلیدی :

rest-fMRI،Self-supervised learning،contrastive learning،CutMix،Continuous Wavelet Transform،SMOTE

چکیده :

Advancements in deep learning have shown substantial promise for medical image analysis, offering potential improvements in healthcare and patient outcomes. However, deep learning models often require large labeled datasets, which are challenging and costly to curate, particularly in the case of fMRI data. Resting-state fMRI (rest-fMRI) presents a unique classification challenge due to its high-dimensional, low sample size nature, making it difficult for traditional deep neural network to achieve reliable accuracy. Self Supervised Learning (SSL), particularly contrastive learning, has emerged as a viable solution to address these limitations. It enables the models to learn meaningful representations unlabeled rest-fMRI data. This study leverages the Continuous Wavelet Transform (CWT) for feature extraction, followed by contrastive learning with CutMix augmentation to capture rich representations from the rest-fMRI time-series data. To address the inherent class imbalance, we apply Synthetic Minority Over-sampling Technique (SMOTE) for data augmentation before final classification. By integrating robust feature extraction, contrastive learning, and targeted data augmentation, our method effectively addresses the challenges posed by high-dimensional data and limited sample sizes. Experimental results demonstrate that our proposed approach achieves high classification accuracy for distinguishing between left versus right epilepsy cases, even with limited and noisy data, while effectively minimizing overfitting.