الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 152 |  |  |
| | | from | 152 | | |
Abstract
Alzheimer’s disease is the most common cause of dementia. It is a significant healthcare burden since it places a considerable strain on caregivers and health- care resources. Early detection and classification of this disease may prevent its progression. Magnetic Resonance Imaging (MRI) is a widely used modality to assess Alzheimer’s disease as it’s non-invasive and uses no radiation, which makes it safe for patients. The problem is that it takes a long time, ranging from 15 minutes up to one hour, to acquire a good-quality image. This causes great dis- comfort to the patients and may prevent its application in time-critical diseases such as strokes.
Compressive sensing is applied to MRI, also known as Compressive Sensing Mag- netic Resonance Imaging (CS-MRI), to overcome the long acquisition time. Com- pressive sensing is a new sensing paradigm that allows sampling a much lower number of data points than required by the well-known Nyquist-Shanon theorem. This reduces the acquisition time significantly; however, it makes both the recon- struction and classification problems more challenging due to the low number of acquired samples.
The reconstruction of the full-sampled data from the compressively acquired, or undersampled, data could be considered the first stage for the high-level objective of disease detection and classification. Many reconstruction approaches have been introduced, including conventional approaches, which suffer from slow iterative procedures and convergence issues. On the other hand, deep learning approaches take a long training time, which is done only once, and a very fast reconstruction that takes a few milliseconds. The reconstructed image is followed by a second stage, which represents a classification network to classify different disease stages. An alternative approach in order to avoid these two sequential stages is to use a compressed learning approach, which is used to directly infer the classes with the help of neural networks without the need to do the reconstruction stage.
In this thesis, a novel conditional Generative Adversarial Network (cGAN) is pro- posed to simultaneously classify and reconstruct the compressively sensed data using a shared network with reduced trainable parameters. The proposed network is called Compressed Learning and Reconstruction GAN (CLRGAN). The CLR- GAN network consists of three main networks: the generator, the critic, and the classifier network. The generator and critic networks are trained in an adversarial framework. In addition to the adversarial loss, a content loss is added to improve the reconstructed image quality. Regarding the classification, it is done directly in the compressed domain and bypasses the reconstruction stage. Furthermore, a sensing network within the CLRGAN is proposed to learn the sensing mask based on the dataset used for training.
The results demonstrate that the proposed CLRGAN is able to learn an under- sampling mask that adapts to the dataset on which it is trained. The CLRGAN achieves very high classification accuracy, reaching 99% using a low sampling rate equal to 0.1, competing with methods that utilize the full sampled data for classifi- cation. Moreover, the proposed reconstruction network achieves comparable image quality to state-of-the-art networks, with PSNR reaching 37 dB and SSIM reach- ing 96.4% using lower trainable parameters. These remarkable results encourage the generalization of the proposed network to assess other diseases.