RT info:eu-repo/semantics/doctoralThesis
T1 DEEP-AD: The deep learning model for diagnostic classification and prognostic prediction of alzheimer's disease
A1 Agarwal, Deevyankar
A2 Universidad de Valladolid. Escuela de Doctorado
K1 Alzheimer, Enfermedad de
K1 Alzheimer's disease
K1 Machine Learning
K1 Aprendizaje profundo
K1 Deep Learning
K1 Alzheimer's
K1 MRI
K1 Resonancia Magnética
K1 33 Ciencias Tecnológicas
AB In terms of context, the aim of this dissertation is to aid neuroradiologists in their clinical judgment regarding the early detection of AD by using DL. To that aim, the system design research methodology is suggested in this dissertation for achieving three goals.The first goal is to investigate the DL models that have performed well at identifying patterns associated with AD, as well as the accuracy so far attained, limitations, and gaps. A systematic review of the literature (SLR) revealed a shortage of empirical studies on the early identification of AD through DL. In this regard, thirteen empirical studies were identified and examined. We concluded that three-dimensional (3D) DL models have been generated far less often and that their performance is also inadequate to qualify them for clinical trials.The second goal is to provide the neuroradiologist with the computer-interpretable information they need to analyze neuroimaging biomarkers. Given this context, the next step in this dissertation is to find the optimum DL model to analyze neuroimaging biomarkers. It has been achieved in two steps. In the first step, eight state-of-the-art DL models have been implemented by training from scratch using end-to-end learning (E2EL) for two binary classification tasks (AD vs. CN and AD vs. stable MCI) and compared by utilizing MRI scans from the publicly accessible datasets of neuroimaging biomarkers. Comparative analysis is carried out by utilizing efficiency-effects graphs, comprehensive indicators, and ranking mechanisms. For the training of the AD vs. sMCI task, the EfficientNet-B0 model gets the highest value for the comprehensive indicator and has the fewest parameters. DenseNet264 performed better than the others in terms of evaluation matrices, but since it has the most parameters, it costs more to train. For the AD vs. CN task by DenseNet264, we achieved 100% accuracy for training and 99.56% accuracy for testing. However, the classification accuracy was still only 82.5% for the AD vs. sMCI task. In the second step, fusion of transfer learning (TL) with E2EL is applied to train the EfficientNet-B0 for the AD vs. sMCI task, which achieved 95.29% accuracy for training and 93.10% accuracy for testing. Additionally, we have also implemented EfficientNet-B0 for the multiclass AD vs. CN vs. sMCI classification task with E2EL to be used in ensemble of models and achieved 85.66% training accuracy and 87.38% testing accuracy.To evaluate the model’s robustness, neuroradiologists must validate the implemented model. As a result, the third goal of this dissertation is to create a tool that neuroradiologists may use at their convenience. To achieve this objective, this dissertation proposes a web-based application (DEEP-AD) that has been created by making an ensemble of Efficient-Net B0 and DenseNet 264 (based on the contribution of goal 2).The accuracy of a DEEP-AD prototype has undergone repeated evaluation and improvement. First, we validated 41 subjects of Spanish MRI datasets (acquired from HT Medica, Madrid, Spain), achieving an accuracy of 82.90%, which was later verified by neuroradiologists. The results of these evaluation studies showed the accomplishment of such goals and relevant directions for future research in applied DL for the early detection of AD in clinical settings.
YR 2023
FD 2023
LK https://uvadoc.uva.es/handle/10324/62581
UL https://uvadoc.uva.es/handle/10324/62581
LA eng
NO Escuela de Doctorado
DS UVaDOC
RD 11-jul-2024