This thesis addressed the challenging problem of predicting the preclinical stage of Alzheimer’s disease (AD) in cognitively normal individuals based on their amyloid deposition levels. Beta-amyloid (Aβ) is a protein that, when abnormally accumulated in the brain, represents one of the earliest and most recognized biomarkers of AD, often preceding the onset of cognitive symptoms. The study relied exclusively on structural magnetic resonance imaging (MRI) scans and focused specifically on the contribution of white matter hyperintensities (WMH) to this task. These lesions, often associated with both vascular and neurodegenerative processes, were placed at the core of our analysis to explore their potential role as early imaging biomarkers of AD. Emerging evidence suggests that WMH burden may interact with amyloid accumulation even in the earliest phases of the disease, raising the question of whether these lesions could serve as supportive imaging markers of preclinical AD. Building on this premise, the main objective of this dissertation was to evaluate the contribution of WMHs to deep learning–based classification of amyloid positivity in cognitively normal individuals. To this end, we trained dedicated Deep Learning (DL) models using FLAIR MRI scans and employed explainable artificial intelligence (XAI) methods to identify the anatomical regions and lesion-related features that most strongly influenced the model’s predictions. To achieve this, we adapted the AXIAL framework proposed by Lozupone et al. (2024), originally designed for AD vs CN classification via 3D MRI analysis using 2D convolutional neural networks (CNNs), and applied it to our amyloid-based prediction task. This framework integrates a soft attention mechanism that allows 2D CNNs to extract volumetric representations while learning the relative importance of each slice, generating voxel-level attention maps. We also implemented Grad-CAM to provide a comparative interpretability baseline. The study was conducted on a dataset from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), comprising 268 subjects, each with a FLAIR MRI volume, a T1-weighted image (used for preprocessing), and a corresponding amyloid deposition value used to distinguish Aβ⁺ (preclinical AD) from Aβ⁻ (healthy control) subjects. WMH masks were extracted via FSL’s BIANCA algorithm and integrated into the explainability analyses to assess their relevance. Our method achieved improved performance compared to similar studies (Hwang et al., 2023) when relying solely on MRI information, reaching an accuracy of 0.870 and a Matthews correlation coefficient (MCC) of 0.725. The performance also remained comparable to the reference AXIAL framework (Lozupone et al., 2024), which reported accuracy = 0.856 and MCC = 0.712, despite our model addressing a subtler task (preclinical AD prediction versus AD vs CN classification). We achieved this prognostic result by adopting a similar procedure to the AXIAL framework and introducing a data augmentation strategy applied to the minority class (Aβ⁺) to balance the training set. This adjustment increased the model’s sensitivity to subtle morphological variations, thereby facilitating the detection of early-stage AD patterns. The explainability analyses, using and comparing attention-based and Grad-CAM approaches, consistently demonstrated that WMHs play a significant role in shaping the model’s decision process when classifying preclinical AD (Aβ⁺) subjects. Supporting the hypothesis that, in the preclinical stage of AD, where overt cortical or subcortical morphological alterations are not yet detectable on MRI, WMHs may constitute one of the few structural cues available to the network. Consequently, as lesion extent increases, WMHs become an increasingly informative and discriminative signal, guiding the model’s predictions toward amyloid positivity. Taken together, these findings highlight the predictive relevance of WMHs in preclinical AD and reinforce the value of incorporating vascular pathology markers alongside traditional AD biomarkers in early diagnostic and risk-stratification frameworks.
Questa tesi ha affrontato il complesso problema della diagnosi della fase preclinica della malattia di Alzheimer (AD) in individui cognitivamente normali, basandosi sui loro livelli di deposizione di amiloide. La beta-amiloide (Aβ) è una proteina che, quando si accumula in modo anomalo nel cervello, rappresenta uno dei biomarcatori più precoci e riconosciuti dell’AD, spesso antecedente alla comparsa dei sintomi cognitivi. Lo studio si è basato esclusivamente su immagini di risonanza magnetica (MRI) strutturale e si è focalizzato in particolare sul contributo delle iperintensità della sostanza bianca (WMH). Queste lesioni, spesso associate a processi sia vascolari sia neurodegenerativi, sono state poste al centro dell’analisi per esplorarne il potenziale ruolo come biomarcatori di imaging precoci dell’AD. Evidenze recenti suggeriscono infatti che il carico di WMH possa interagire con l’accumulo di amiloide già nelle primissime fasi della malattia, sollevando la questione se tali lesioni possano fungere da marcatori di imaging di supporto nello stadio preclinico. Partendo da questo presupposto, l’obiettivo principale di questa tesi è stato valutare il contributo delle WMH nella classificazione, mediante deep learning (DL), della positività amiloide in individui cognitivamente normali. A tal fine, sono stati addestrati modelli di DL dedicati utilizzando immagini FLAIR e sono state impiegate tecniche di intelligenza artificiale spiegabile (XAI) per identificare le regioni anatomiche e le caratteristiche delle lesioni che influenzano maggiormente le predizioni del modello. Per raggiungere questo obiettivo, è stato adattato il framework AXIAL proposto da Lozupone et al. (2024), originariamente sviluppato per la classificazione AD vs CN tramite analisi MRI 3D basata su reti neurali convoluzionali 2D (CNN), applicandolo al nostro compito di previsione basato sull’amiloide. Questo framework integra un meccanismo di soft attention che consente alle CNN 2D di estrarre rappresentazioni volumetriche apprendendo l’importanza relativa di ciascuna slice, generando mappe di attenzione a livello voxel. È stato inoltre implementato il metodo Grad-CAM per fornire un confronto interpretativo parallelo. Lo studio è stato condotto su un dataset dell’Alzheimer’s Disease Neuroimaging Initiative (ADNI), composto da 268 soggetti, ciascuno con un volume MRI FLAIR, un’immagine T1 (utilizzata per il preprocessing) e un valore di deposizione amiloide impiegato per distinguere i soggetti Aβ⁺ (AD preclinico) dagli Aβ⁻ (controlli sani). Le maschere WMH sono state estratte tramite l’algoritmo BIANCA di FSL e integrate nell’analisi di explainability per valutarne la rilevanza. Il nostro metodo ha ottenuto prestazioni superiori a studi analoghi (Hwang et al., 2023) basati unicamente su MRI, raggiungendo un’accuratezza di 0.870 e un coefficiente MCC di 0.725. Le performance sono risultate comparabili al framework AXIAL di riferimento (Lozupone et al., 2024), che riporta accuratezza = 0.856 e MCC = 0.712, nonostante il nostro modello affronti un compito più sottile (previsione dello stadio preclinico rispetto alla classificazione AD vs CN). Questo risultato è stato ottenuto adottando una procedura ispirata al framework AXIAL e introducendo una strategia di data augmentation applicata alla classe minoritaria (Aβ⁺) per bilanciare il set di training. Tale accorgimento ha aumentato la sensibilità del modello alle variazioni morfologiche più sottili, facilitando la rilevazione dei pattern dell’AD precoce. Le analisi di explainability, basate sul confronto tra approcci attention-based e Grad-CAM, hanno dimostrato in modo coerente che le WMH svolgono un ruolo significativo nel processo decisionale del modello nella classificazione dei soggetti Aβ⁺. Ciò supporta l’ipotesi che, nello stadio preclinico dell’AD, in cui le alterazioni morfologiche corticali o sottocorticali non sono ancora visibili alla MRI, le WMH possano rappresentare uno dei pochi indizi strutturali disponibili al modello. Di conseguenza, all’aumentare dell’estensione delle lesioni, le WMH diventano un segnale sempre più informativo e discriminativo, guidando le predizioni verso la positività amiloide. Complessivamente, questi risultati evidenziano la rilevanza predittiva delle WMH nella fase preclinica dell’AD e rafforzano il valore dell’integrazione di indicatori di compromissione vascolare insieme ai biomarcatori tradizionali dell’AD nei framework diagnostici e di stratificazione del rischio precoce.
Understanding the role of white matter hyperintensities in preclinical Alzheimer's disease: an explainable AI investigation
Tomasella, Andrea
2024/2025
Abstract
This thesis addressed the challenging problem of predicting the preclinical stage of Alzheimer’s disease (AD) in cognitively normal individuals based on their amyloid deposition levels. Beta-amyloid (Aβ) is a protein that, when abnormally accumulated in the brain, represents one of the earliest and most recognized biomarkers of AD, often preceding the onset of cognitive symptoms. The study relied exclusively on structural magnetic resonance imaging (MRI) scans and focused specifically on the contribution of white matter hyperintensities (WMH) to this task. These lesions, often associated with both vascular and neurodegenerative processes, were placed at the core of our analysis to explore their potential role as early imaging biomarkers of AD. Emerging evidence suggests that WMH burden may interact with amyloid accumulation even in the earliest phases of the disease, raising the question of whether these lesions could serve as supportive imaging markers of preclinical AD. Building on this premise, the main objective of this dissertation was to evaluate the contribution of WMHs to deep learning–based classification of amyloid positivity in cognitively normal individuals. To this end, we trained dedicated Deep Learning (DL) models using FLAIR MRI scans and employed explainable artificial intelligence (XAI) methods to identify the anatomical regions and lesion-related features that most strongly influenced the model’s predictions. To achieve this, we adapted the AXIAL framework proposed by Lozupone et al. (2024), originally designed for AD vs CN classification via 3D MRI analysis using 2D convolutional neural networks (CNNs), and applied it to our amyloid-based prediction task. This framework integrates a soft attention mechanism that allows 2D CNNs to extract volumetric representations while learning the relative importance of each slice, generating voxel-level attention maps. We also implemented Grad-CAM to provide a comparative interpretability baseline. The study was conducted on a dataset from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), comprising 268 subjects, each with a FLAIR MRI volume, a T1-weighted image (used for preprocessing), and a corresponding amyloid deposition value used to distinguish Aβ⁺ (preclinical AD) from Aβ⁻ (healthy control) subjects. WMH masks were extracted via FSL’s BIANCA algorithm and integrated into the explainability analyses to assess their relevance. Our method achieved improved performance compared to similar studies (Hwang et al., 2023) when relying solely on MRI information, reaching an accuracy of 0.870 and a Matthews correlation coefficient (MCC) of 0.725. The performance also remained comparable to the reference AXIAL framework (Lozupone et al., 2024), which reported accuracy = 0.856 and MCC = 0.712, despite our model addressing a subtler task (preclinical AD prediction versus AD vs CN classification). We achieved this prognostic result by adopting a similar procedure to the AXIAL framework and introducing a data augmentation strategy applied to the minority class (Aβ⁺) to balance the training set. This adjustment increased the model’s sensitivity to subtle morphological variations, thereby facilitating the detection of early-stage AD patterns. The explainability analyses, using and comparing attention-based and Grad-CAM approaches, consistently demonstrated that WMHs play a significant role in shaping the model’s decision process when classifying preclinical AD (Aβ⁺) subjects. Supporting the hypothesis that, in the preclinical stage of AD, where overt cortical or subcortical morphological alterations are not yet detectable on MRI, WMHs may constitute one of the few structural cues available to the network. Consequently, as lesion extent increases, WMHs become an increasingly informative and discriminative signal, guiding the model’s predictions toward amyloid positivity. Taken together, these findings highlight the predictive relevance of WMHs in preclinical AD and reinforce the value of incorporating vascular pathology markers alongside traditional AD biomarkers in early diagnostic and risk-stratification frameworks.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Tomasella_Tesi.pdf
solo utenti autorizzati a partire dal 19/11/2026
Descrizione: Contenuto Tesi
Dimensione
13.41 MB
Formato
Adobe PDF
|
13.41 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Tomasella_Executive Summary.pdf
solo utenti autorizzati a partire dal 19/11/2026
Descrizione: Contenuto Excecutive Summary
Dimensione
1.2 MB
Formato
Adobe PDF
|
1.2 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/10589/247102