This paper aims at implementing novel biomarkers extracted from functional magnetic resonance imaging (fMRI) images taken at resting-state using convolutional neural networks (CNN) to predict relapse in heavy smoker subjects. In this regard, two classes of subjects were studied. The first class contains 19 subjects that took the drug N-acetylcysteine (NAC), and the second class contains 20 subjects that took a placebo. The subjects underwent a double-blind smoking cessation treatment. The resting-state fMRI of the subjects' brains were recorded through 200 snapshots before and after the treatment. The relapse data was assessed after 6 months past the treatment. The data was pre-processed and an undercomplete autoencoder along with various similarity metrics was developed to extract salient features that could differentiate the pre and post treatment images. Finally, the extracted feature matrix was fed into robust classification algorithms to classify the subjects in terms of relapse and non-relapse. The XGBoost algorithm with 0.86 precision and an AUC of 0.92 outperformed the other classification methods in prediction of relapse in subjects.