Nowadays, Alzheimer’s disease is becoming a major public health issue worldwide. Thus, a higher knowledge yet early diagnosis of it may be fundamental to both slow down the development of symptoms and enable better therapeutic interventions. Clinically, neuroimaging techniques such as magnetic resonance imaging are available for the Alzheimer’s disease diagnosis. Through magnetic resonance imaging, volumetric scans can be obtained, helping in the detection of structural abnormalities and tracking of the evolution of brain atrophy. Deep learning algorithms for Alzheimer’s disease diagnosis applied to volumetric scans are increasingly used in medical field, but they are still not trusted by clinicians because they lack interpretability. This thesis has a dual purpose. Firstly, it has been performed a review of state-of-the-art studies which applied interpretability algorithms for Alzheimer’s disease diagnosis in order to understand the current trends. Then, it has been conducted an analysis of volumetric magnetic resonance scans by exploiting two convolutional neural networks and comparing their performance: a pre-trained 3D convolutional neural network (C3DKeras) and an end-to-end time-distributed one. To fulfil the first task, a descriptive literature review has been performed, whereas for the second one, a Python-based implementation was conducted. According to the literature outcomes, there is still uncertainty concerning the best interpretability technique to be applied for Alzheimer’s disease diagnosis, even though attribution map approaches seem to produce the most coherent interpretations. For what concerns convolutional neural networks for volumetric data processing, the end-to-end time-distributed one resulted to be the best approach because of its higher performance and lower computational cost. A future development of this thesis could be the addition of an interpretability module to the end-to-end time-distributed convolutional neural network in order to make a step forward in the direction of an interpretable Alzheimer’s disease diagnosis.

Nowadays, Alzheimer’s disease is becoming a major public health issue worldwide. Thus, a higher knowledge yet early diagnosis of it may be fundamental to both slow down the development of symptoms and enable better therapeutic interventions. Clinically, neuroimaging techniques such as magnetic resonance imaging are available for the Alzheimer’s disease diagnosis. Through magnetic resonance imaging, volumetric scans can be obtained, helping in the detection of structural abnormalities and tracking of the evolution of brain atrophy. Deep learning algorithms for Alzheimer’s disease diagnosis applied to volumetric scans are increasingly used in medical field, but they are still not trusted by clinicians because they lack interpretability. This thesis has a dual purpose. Firstly, it has been performed a review of state-of-the-art studies which applied interpretability algorithms for Alzheimer’s disease diagnosis in order to understand the current trends. Then, it has been conducted an analysis of volumetric magnetic resonance scans by exploiting two convolutional neural networks and comparing their performance: a pre-trained 3D convolutional neural network (C3DKeras) and an end-to-end time-distributed one. To fulfil the first task, a descriptive literature review has been performed, whereas for the second one, a Python-based implementation was conducted. According to the literature outcomes, there is still uncertainty concerning the best interpretability technique to be applied for Alzheimer’s disease diagnosis, even though attribution map approaches seem to produce the most coherent interpretations. For what concerns convolutional neural networks for volumetric data processing, the end-to-end time-distributed one resulted to be the best approach because of its higher performance and lower computational cost. A future development of this thesis could be the addition of an interpretability module to the end-to-end time-distributed convolutional neural network in order to make a step forward in the direction of an interpretable Alzheimer’s disease diagnosis.

Alzheimer’s disease early diagnosis using convolutional neural networks from volumetric scans

BERNARDI, GIADA
2021/2022

Abstract

Nowadays, Alzheimer’s disease is becoming a major public health issue worldwide. Thus, a higher knowledge yet early diagnosis of it may be fundamental to both slow down the development of symptoms and enable better therapeutic interventions. Clinically, neuroimaging techniques such as magnetic resonance imaging are available for the Alzheimer’s disease diagnosis. Through magnetic resonance imaging, volumetric scans can be obtained, helping in the detection of structural abnormalities and tracking of the evolution of brain atrophy. Deep learning algorithms for Alzheimer’s disease diagnosis applied to volumetric scans are increasingly used in medical field, but they are still not trusted by clinicians because they lack interpretability. This thesis has a dual purpose. Firstly, it has been performed a review of state-of-the-art studies which applied interpretability algorithms for Alzheimer’s disease diagnosis in order to understand the current trends. Then, it has been conducted an analysis of volumetric magnetic resonance scans by exploiting two convolutional neural networks and comparing their performance: a pre-trained 3D convolutional neural network (C3DKeras) and an end-to-end time-distributed one. To fulfil the first task, a descriptive literature review has been performed, whereas for the second one, a Python-based implementation was conducted. According to the literature outcomes, there is still uncertainty concerning the best interpretability technique to be applied for Alzheimer’s disease diagnosis, even though attribution map approaches seem to produce the most coherent interpretations. For what concerns convolutional neural networks for volumetric data processing, the end-to-end time-distributed one resulted to be the best approach because of its higher performance and lower computational cost. A future development of this thesis could be the addition of an interpretability module to the end-to-end time-distributed convolutional neural network in order to make a step forward in the direction of an interpretable Alzheimer’s disease diagnosis.
2021
2022-07-18
Alzheimer’s disease early diagnosis using convolutional neural networks from volumetric scans
Nowadays, Alzheimer’s disease is becoming a major public health issue worldwide. Thus, a higher knowledge yet early diagnosis of it may be fundamental to both slow down the development of symptoms and enable better therapeutic interventions. Clinically, neuroimaging techniques such as magnetic resonance imaging are available for the Alzheimer’s disease diagnosis. Through magnetic resonance imaging, volumetric scans can be obtained, helping in the detection of structural abnormalities and tracking of the evolution of brain atrophy. Deep learning algorithms for Alzheimer’s disease diagnosis applied to volumetric scans are increasingly used in medical field, but they are still not trusted by clinicians because they lack interpretability. This thesis has a dual purpose. Firstly, it has been performed a review of state-of-the-art studies which applied interpretability algorithms for Alzheimer’s disease diagnosis in order to understand the current trends. Then, it has been conducted an analysis of volumetric magnetic resonance scans by exploiting two convolutional neural networks and comparing their performance: a pre-trained 3D convolutional neural network (C3DKeras) and an end-to-end time-distributed one. To fulfil the first task, a descriptive literature review has been performed, whereas for the second one, a Python-based implementation was conducted. According to the literature outcomes, there is still uncertainty concerning the best interpretability technique to be applied for Alzheimer’s disease diagnosis, even though attribution map approaches seem to produce the most coherent interpretations. For what concerns convolutional neural networks for volumetric data processing, the end-to-end time-distributed one resulted to be the best approach because of its higher performance and lower computational cost. A future development of this thesis could be the addition of an interpretability module to the end-to-end time-distributed convolutional neural network in order to make a step forward in the direction of an interpretable Alzheimer’s disease diagnosis.
File in questo prodotto:
File Dimensione Formato  
BERNARDI_GIADA_TESI.pdf

embargo fino al 17/07/2024

Descrizione: TESI GIADA BERNARDI
Dimensione 4.15 MB
Formato Adobe PDF
4.15 MB Adobe PDF

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12075/9421