The aim of this work is to predict the progression of Alzheimer’s Disease (AD) in Structural Magnetic Resonance Imaging (sMRI) using Deep Generative methods. To the best of our knowledge, this is the first attempt to generate this progression using Deep Learning methods. Alzheimer’s disease is the most common cause of dementia worldwide, and this tendency is predicted to become even more marked in the next years, due the global aging of the population. While several therapies are currently being studied, they are mostly applied when patients experience the first symptoms of cognitive impairment, indicating that the disease is already in an advanced stage. A robust model able to predict the development of the disease and its influence on specific regions of the brain would guarantee higher chances to slow down, stop or even prevent the disease. We use sMRI as the input data for this study for its being relatively cheap and non-invasive for the patient. We focus on the key regions for AD, namely hippocampus and ventricles, by extracting slices as well as 3D organs using a Fully Convolutional segmentation network. Both the 2D slices and the 3D shapes are then analysed using Convolutional Variational Autoencoders (CVAE) and Conditional Adversarial Autoencoders (CAA), integrating both supervised and unsupervised approaches. The Convolutional Variational Autoencoder is used to learn a manifold representation that encodes the most distinctive brain features and can be walked to progress them in terms of shape, size and morphological characteristics. The Conditional Adversarial Autoencoder, an integration of Autoencoders and Generative Adversarial Networks, is used to generate a progression in time of the input brain volumes. This progression is evaluated qualitatively on 2D slices and quantitatively on 3D shapes, in the latter case showing a statistically significant decrease in the shape of the hippocampus over time, that is more evident in AD subjects as opposed to NC, as confirmed by literature.
Obiettivo di questo lavoro di tesi è di prevedere la progressione della malattia di Alzheimer (AD) in immagini di risonanza magnetica strutturale (sMRI) utilizzando modelli generativi deep. Al meglio della nostra conoscenza, questo è il primo tentativo di generare una progressione della malattia utilizzando metodi di Deep Learning. La malattia di Alzheimer è la più comune causa di demenza, una tendenza destinata a diventare sempre più marcata nei prossimi anni con l’invecchiamento globale della popolazione. Diverse terapie sono attualmente in corso di sperimentazione, ma sono quasi sempre applicate a pazienti che mostrano già i primi sintomi di declino cognitivo, ovvero nei quali la malattia è ormai presente in stato avanzato. Un modello robusto in grado di predire lo sviluppo della malattia e la sua influenza su regioni specifiche del cervello potrebbe permettere di rallentare, fermare o prevenire la malattia. La risonanza magnetica strutturale è stata utilizzata per questo studio perché meno costosa e meno invasiva per il paziente rispetto ad altri esami. Questo studio si focalizza sulle regioni chiave per l’Alzheimer, come l’ippocampo e i ventricoli, estraendo slice e organi 3D utilizzando una rete di segmentazione Fully Convolutional. Le slice 2D e gli organi 3D sono poi analizzati utilizzando un Autoencoder Variazionale Convoluzionale (CVAE) e un Autoencoder Condizionale Avversario (CAAE), integrando un approccio supervisionato ad uno non supervisionato. Il CVAE è utilizzato per imparare un manifold in grado di codificare le feature più rilevanti per la diagnosi della patologie, ed essere attraversato per ottenere una progressione in forma, volume e caratteristiche morfologiche. Il CAAE, un’integrazione di Autoencoders e Reti Generative Avversarie, è utilizzato per generare una progressione nel tempo dei volumi in input. Questa progressione è valutata qualitativamente sulle slice 2D e quantitativamente sugli organi 3D. Nel secondo caso, si evidenzia una decrescita statisticamente significativa nella dimensione dell’ippocampo nel tempo, che è più evidente nei soggetti AD rispetto ai soggetti di controllo, come confermato dalla letteratura.
Deep generative models for predicting Alzheimer's disease progression from MR data
MILANA, DILETTA
2016/2017
Abstract
The aim of this work is to predict the progression of Alzheimer’s Disease (AD) in Structural Magnetic Resonance Imaging (sMRI) using Deep Generative methods. To the best of our knowledge, this is the first attempt to generate this progression using Deep Learning methods. Alzheimer’s disease is the most common cause of dementia worldwide, and this tendency is predicted to become even more marked in the next years, due the global aging of the population. While several therapies are currently being studied, they are mostly applied when patients experience the first symptoms of cognitive impairment, indicating that the disease is already in an advanced stage. A robust model able to predict the development of the disease and its influence on specific regions of the brain would guarantee higher chances to slow down, stop or even prevent the disease. We use sMRI as the input data for this study for its being relatively cheap and non-invasive for the patient. We focus on the key regions for AD, namely hippocampus and ventricles, by extracting slices as well as 3D organs using a Fully Convolutional segmentation network. Both the 2D slices and the 3D shapes are then analysed using Convolutional Variational Autoencoders (CVAE) and Conditional Adversarial Autoencoders (CAA), integrating both supervised and unsupervised approaches. The Convolutional Variational Autoencoder is used to learn a manifold representation that encodes the most distinctive brain features and can be walked to progress them in terms of shape, size and morphological characteristics. The Conditional Adversarial Autoencoder, an integration of Autoencoders and Generative Adversarial Networks, is used to generate a progression in time of the input brain volumes. This progression is evaluated qualitatively on 2D slices and quantitatively on 3D shapes, in the latter case showing a statistically significant decrease in the shape of the hippocampus over time, that is more evident in AD subjects as opposed to NC, as confirmed by literature.File | Dimensione | Formato | |
---|---|---|---|
tesi.pdf
accessibile in internet per tutti
Descrizione: Thesis - update 2
Dimensione
28.53 MB
Formato
Adobe PDF
|
28.53 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/137726