Robustness of quantitative imaging biomarkers for classification of neurodegenerative disorders

Štokelj, Eva (author), Simončič, Urban (mentor)

This thesis evaluates the robustness of quantitative imaging biomarkers (QIBs) for the classification of neurodegenerative diseases. Although many biomarkers distinguish patients from healthy controls at the group level, their clinical usefulness depends on reliable performance under biological and methodological variability. A unified robustness framework was developed, based on effect size (the absolute value of Cohen's d, |d|), its relationship to classification performance (AUC), and systematic sensitivity analysis. Three categories of biomarkers were investigated. The first comprised metabolic brain patterns derived from FDG-PET using the scaled subprofile model/principal component analysis (SSM/PCA). These showed excellent diagnostic performance for Alzheimer's and Parkinson's disease (AUC ≈ 0.9; |d| ≈ 1.7–1.8). A novel multi-group extension was developed to simultaneously differentiate Parkinson's disease, multiple system atrophy, and progressive supranuclear palsy, achieving large effect sizes at the reference configuration (AUC ≈ 0.9–1.0; |d| ≈ 2.1–2.4) without requiring healthy controls. The second category, the asymmetry index from structural MRI, detected differences between Alzheimer's patients and controls, but diagnostic performance remained moderate (AUC ≈ 0.6; |d| ≈ 0.5) and did not improve across methodological variations, indicating limited potential within the tested implementation. The third category comprised diffusion-MRI biomarkers. TBSS and network-based tractography combined diagnostic performance and robustness more favorably (AUC ≈ 0.7–0.8; |d| ≈ 0.9–1.0) than histogram-based approaches (AUC ≈ 0.7; |d| ≈ 0.7), though they were more sensitive to methodological choices than metabolic patterns. Robustness and discriminative strength do not necessarily coincide: a biomarker may be robust yet weakly discriminative, or strongly discriminative yet not robust, so the two must be assessed jointly. Among the approaches studied, metabolic brain patterns were the most robust and clinically promising. The proposed approach enables a realistic comparison of QIBs under biological and methodological variability.

• Language: English
• Keywords: Quantitative imaging biomarkers, neurodegenerative diseases, Alzheimer’s disease, Parkinson’s disease, parkinsonisms, FDG-PET, MRI, DTI, disease-related metabolic patterns, asymmetry index, diffusion measures from MR images
• Work type: Doctoral dissertation
• Typology: 2.08 - Doctoral Dissertation
• Organization: FMF - Faculty of Mathematics and Physics
• Year: 2026
• PID: 20.500.12556/RUL-190591
• COBISS.SI-ID: 303897347
• Publication date in RUL: 30.10.2026

Secondary language

• Language: Slovenian
• Title: Robustnost kvantitativnih slikovnih bioloških označevalcev za razvrščanje nevrodegenerativnih bolezni

Abstract

Doktorska naloga obravnava robustnost kvantitativnih slikovnih bioloških označevalcev (QIB) za razvrščanje nevrodegenerativnih bolezni. Čeprav številni označevalci razlikujejo med bolniki in zdravimi posamezniki na ravni skupin, njihova klinična uporabnost zahteva zanesljivo delovanje ob biološki in metodološki variabilnosti. Razvit je bil enoten okvir vrednotenja robustnosti, ki temelji na velikosti učinka (absolutna vrednost Cohenovega d, |d|), njeni povezavi s klasifikacijsko uspešnostjo (AUC) in sistematični analizi občutljivosti. Preučene so bile tri skupine označevalcev. Prvo predstavljajo presnovni možganski vzorci, pridobljeni iz FDG-PET slik z metodo SSM/PCA. Ti so dosegli zelo visoko diagnostično uspešnost pri Alzheimerjevi in Parkinsonovi bolezni (AUC ≈ 0,9; |d| ≈ 1,7–1,8). Razvita je bila novost — več-skupinska razširitev za sočasno razlikovanje med Parkinsonovo boleznijo, multiplo sistemsko atrofijo in progresivno supranuklearno paralizo, ki je pri referenčni konfiguraciji dosegla visoke velikosti učinka (AUC ≈ 0,9–1,0; |d| ≈ 2,1–2,4) brez potrebe po zdravih kontrolah. Druga skupina, indeks asimetrije iz strukturnih MR slik, je zaznala razlike med bolniki z Alzheimerjevo boleznijo in zdravimi, a je bila diagnostična uspešnost le zmerna (AUC ≈ 0,6; |d| ≈ 0,5) in se z metodološkimi spremembami ni izboljšala, kar kaže na omejen potencial v okviru uporabljene implementacije. Tretjo skupino predstavljajo označevalci iz difuzijskega MRI. TBSS in omrežna traktografija sta združila diagnostično uspešnost in robustnost ugodneje (AUC ≈ 0,7–0,8; |d| ≈ 0,9–1,0) kot histogramski pristopi (AUC ≈ 0,7; |d| ≈ 0,7), a sta bila bolj občutljiva na metodološke izbire kot presnovni vzorci. Robustnost in diagnostična uspešnost se ne ujemata nujno: označevalec je lahko robusten, a šibko ločevalen, ali močno ločevalen, a nerobusten, zato je treba obe lastnosti ocenjevati skupaj. Med preučenimi pristopi so bili presnovni možganski vzorci najbolj robustni in klinično obetavni. Predlagani okvir omogoča realistično primerjavo označevalcev ob biološki in metodološki variabilnosti.

• Keywords: Kvantitativni slikovni biološki označevalci, nevrodegenerativne bolezni, Alzheimerjeva bolezen, Parkinsonova bolezen, parkinsonizmi, FDG-PET, MRI, DTI, Presnovni možganski vzorci, indeks asimetrije, difuzijske mere iz MR slik