Medical Image Reconstruction Using Generative Adversarial Network for Alzheimer Disease Assessment with Class-Imbalance Problem

摘要

One of the most challenging problem faced with medical image analysis is the lack of some modality images. In this work, we propose an effective data augmentation method that uses the generative adversarial network to reconstruct the missing PET images. A densely connected convolutional network is developed as the classification model to make the binary classification. The experiments on ADNI class-imbalanced dataset demonstrate that add the reconstructed images can significantly improve the classification performance of the densely connected model and effectively deal with the class-imbalanced challenge. The influence of different noisy dimensions is also detailedly discussed in term of maximum mean discrepancy and structural similarity metric. The proposed method will make some contribution to other clinical class-imbalanced datasets.