A new deep learning-based method for data augmentation: application to Alzheimer dementia diagnosis.

Abstract

Deep learning has demonstrated remarkable capabilities in the field of medical imaging, one of the significant advantages of deep learning in medical imaging is its ability to automate and enhance various processes. For example, deep learning models can automatically detect abnormalities in medical images, and assist in tumor segmentation. However, the availability of medical imaging data can be a significant challenge in developing deep learning models. Col lecting large amounts of high-quality annotated medical imaging data can be time-consuming, expensive, and sometimes limited due to privacy concerns and data-sharing regulations. By utilizing few-shot learning, researchers and medical professionals can work with smaller datasets. The objective of this thesis is to advance the field of few-shot learning by focusing on the data perspective. In order to achieve this goal, we developed a novel architecture based on delta autoencoder, Our proposed architecture focuses on learning the transformation between different classes, enabling the generation of data that exhibits characteristics of multiple classes. To evaluate the effectiveness of the proposed architecture, the OASIS 1 dataset was uti lized as a benchmark. The experiments conducted on this dataset demonstrated significant improvements in few-shot learning performance compared to existing methods. Specifically, the achieved result of 82% accuracy showcases the remarkable capabilities of the developed architecture in enhancing the learning process. The findings showcase the potential of the proposed method in improving generalization and adaptability when faced with limited labeled data. These results encourage further exploration and development of innovative approaches to enhance machine learning algorithms in few-shot learning scenarios.