TA-GAN: transformer-driven addiction-perception generative adversarial network

Changhong Jing, Changwei Gong, Zuxin Chen, Baiying Lei, Shuqiang Wang

Abstract: The identification of addiction-related brain connections using functional magnetic resonance imaging (fMRI) is essential for comprehending the mechanisms of addiction. However, it is a challenge to effectively identify addiction-related brain connections using fMRI for traditional methods. In this work, the transformer-driven addiction-perception generative adversarial network (TA-GAN) is proposed to identify brain connectivity associated with nicotine addiction. In particular, the generator of TA-GAN takes into account that the convolutional neural network (CNN) can capture the local spatial features between brain regions, while the transformer specializes in extracting global brain connectivity information. Specifically, the external encoder-decoder structure aims to extract and reconstruct representations of brain region features. The transformer structure is implemented to extract global dependencies between brain region features. The discriminator is frequently overfitting when Generative Adversarial Networks (GANs) are trained with insufficient data. We proposed an adaptive discriminator enhancement mechanism that allows the discriminator to acquire addiction-related brain connections with limited data volume efficiently. Validation results on rat nicotine addiction data show that our proposed method achieves promising results in both qualitative and quantitative measurements.