An Industrial-Grade Brain Imaging Based Deep Learning Classifier

National Natural Science Foundation of China (NSFC)

PI：Prof. Chao-Gan Yan

Magnetic resonance structural brain image-based deep learning algorithms have achieved good performance on predicting chronological age of human being, but few researches look into the performance of deep convolutional neural network on predicting biological gender of human. In this study, we pooled more than 34 datasets and constituted the biggest brain magnetic resonance image sample (n = 85,721) to date. And we applied state of the art architecture of deep convolutional neural network, Inception-ResNet-V2, on the pooled data. The performance of gender classifier reached averaging 94.4% accuracy in leave-sites-out 5-fold cross-validation. To explore the potential of deep convolutional network on objective diagnose for brain related disorders, we tested the performance of the fine-tuned classifier on disease samples through a transfer learning framework. The results showed that transferred Alzheimer’s disease and attention deficit hyperactivity disorder disease classifier achieved satisfying accuracy (89% and 64%) but the others kept chance-level accuracy in leave-sites-out cross-validation.