Early diagnosis of Alzheimer's disease from functional rs-fMRI images based on deep learning networks and transfer...

Exploiting deep learning methods to accelerate the analysis of medical images and the interpretation of pathology results for early diagnosis of Alzheimer's disease (AD) has recently attracted great attention. However, challenges like sub-optimal classifiers and poor image representation hinder their effectiveness. Computer-aided diagnosis (CADx) can improve performance by classifying patterns. Despite the drawbacks of deep networks such as Visual Geometric Group (VGG), including long processing times and performance issues due to data distribution, many CADx systems still rely on VGG classifiers due to their potential for high accuracy when properly trained. To tackle these issues, this paper introduces two novel deep networks, called optimized VGG-16 (OVGG-16) and optimized VGG-19 (OVGG-19), in light of the concepts of transfer learning and dense layers to improve diagnosis performance. The proposed system was developed for the diagnosis of AD employing the OVGG-16 and OVGG-19 networks as classifiers from rs-fMRI images. The results show that the convergence rate of the proposed OVGG-16 and OVGG-19 networks is more rapid than that of the conventional VGG-16 and VGG-19. Moreover, the proposed system, which uses the OVGG-16 network, yielded a high accuracy of 100% and 98.83% for binary and multiclass classification, respectively, which surpasses existing state-of-the-art approaches.