Identifikacija značilnih presnovnih možganskih vzorcev pogostih demenc in njihova uporaba v računalniško podprtih diferencialno diagnostičnih orodjih

Perovnik, Matej (author), Trošt, Maja (mentor), Jeraj, Robert (comentor)

Ozadje: V najzgodnejših stopnjah nevrodegenerativnih demenc lahko diferencialna diagnoza predstavlja izziv, zaradi česar postaja uporaba bioloških označevalcev v klinični praksi in pri raziskovanju vedno pogostejša. Izboljšava trenutnih in razvoj novih bioloških označevalcev sta prioriteti v nevroznanosti in klinični nevrologiji za izboljšanje diagnostične natančnosti. Funkcijsko slikanje možganov z 2-[18F]fluoro-2-deoksi-D-glukozo in pozitronsko izsevno tomografijo (angl. 2-[18F]fluoro-2-deoxy-D-glucose positron emission tomography, 2-[18F]FDG PET) lahko zazna spremembe v presnovi možganov, ki odražajo potekajoč nevrodegenerativni proces. Uporaba metod za mrežno analizo na slikah 2-[18F]FDG PET nam omogoča identifikacijo značilnih z boleznijo povezanih presnovnih možganskih vzorcev. Te vzorce lahko kvantificiramo pri posameznem bolniku in njihovo izraženost uporabimo pri diferencialni diagnozi, prognozi, sledenju poteka bolezni in, sčasoma tudi, odzivu na zdravljenje, ki spreminja potek bolezni. Cilji: identifikacija in validacija značilnih presnovnih možganskih vzorcev dveh najpogostejših nevrodegenerativnih demenc: demence zaradi Alzheimerjeve bolezni (angl. dementia due to Alzheiemer's disease, AD) in demence z Lewyjevimi telesci (angl. dementia with Lewy bodies, DLB) z uporabo mrežne analize in slik 2-[18F]FDG PET, razvoj avtomatiziranega klasifikacijska algoritma za diferenciacijo in klasifikacijo slik 2-[18F]FDG PET bolnikov z različnimi demencami ter primerjava njegove diagnostične vrednosti z vidnimi odčitki strokovnjakov. Metode: Analizirali smo klinične podatke in slike 2-[18F]FDG PET 388 preiskovancev: 133 bolnikov z AD, 42 bolnikov z blago kognitivno motnjo (angl. mild cognitive impairment, MCI), 79 bolnikov z DLB, 23 bolnikov z vedenjsko obliko frontotemporalne demence (angl. behavioural variant frontotemporal dementia, bvFTD) in 111 zdravih kontrol (angl. normal controls, NC). Po predpripravi slik smo (1.) uporabili metodo skaliranega subprofilnega modela, ki temelji na analizi glavnih komponent (angl. scaled subprofile model/principal component analysis, SSM/PCA), na slikah 2-[18F]FDG PET 20 bolnikov z AD in 20 NC za identifikacijo značilnega presnovnega vzorca pri AD (angl. AD-related pattern, ADRP). Izračunali smo diagnostično natančnost vzorca in ga validirali na treh neodvisnih kohortah bolnikov. Dodatno smo izračunali tudi korelacijo med izraženostjo vzorca in bolnikovo oceno splošnega kognitivnega stanja ter trajanjem bolezni. Prav tako smo (2.) metodo SSM/PCA uporabili na slikah 2-[18F]FDG PET 20 bolnikov z DLB in 20 NC za identifikacijo značilnega presnovnega vzorca pri DLB (angl. DLB-related pattern, DLBRP). Izračunali smo diagnostično natančnost vzorca in ga validirali na neodvisni validacijski kohorti, ki je bila sestavljena iz 59 bolnikov z DLB, 63 bolnikov z AD in 21 NC. Dodatno smo izračunali korelacijo med izraženostjo vzorca in bolnikovimi ocenami splošnega kognitivnega stanja ter trajanjem bolezni. Prav tako smo (3.) izgradili in testirali model strojnega učenja, osnovan na dveh naborih značilnosti (izraženosti značilnih z boleznijo povezanih vzorcev in povprečnem privzemu glukoze v 95 področjih zanimanja (angl. regions of interest, ROIs)). Izračunali smo diagnostično vrednost klasifikacijskega algoritma in jo primerjali z vidno oceno strokovnjakov. Zlati diagnostični standard sta bila klinična diagnoza na kontrolnem pregledu (M = 25 ± 20 mesecev) oziroma izvid likvorskih bioloških označevalcev za Alzheimerjevo bolezen. Rezultati: (1.) Za novoodkriti ADRP so bila značilna področja relativno znižane presnovne aktivnosti v temporoparietalni skorji, posteriornem cingulumu in prekuneusu, ki so bila soodvisna s področji relativno zvišane presnovno aktivnosti v malih možganih. Izraženost ADRP je pomembno razlikovala med AD in NC (angl. areas under the curve (AUCs) = 0,95–0,98) in je korelirala z merami splošnega kognitivnega stanja v vseh vključenih kohortah bolnikov z AD. (2.) Za novoodkriti DLBRP so bila značilna področja znižane presnovne aktivnosti v okcipitalnem, inferiornem parietalnem in inferiornem temporalnem režnju ter prekuneusu obojestransko, ki so bila soodvisna s področji relativno zvišane presnovne aktivnosti v bazalnih ganglijih in srednji temporalni skorji. Izraženost DLBRP je pomembno razlikovala med DLB in NC (AUC = 0,99) in je korelirala z merami splošnega kognitivnega stanja. Po korekciji za topografsko prekrivanje med ADRP in DLBRP je vzorec pomembno razlikoval tudi med DLB in AD (AUC = 0,87). (3.) Klasifikatorja, ki sta temeljila na izraženosti presnovnih vzorcev in privzemu glukoze v ROI, sta dosegla visoko diagnostično natančnost z visoko specifičnostjo in občutljivostjo v vseh štirih diagnostičnih kategorijah (AD, DLB, bvFTD in NC). Oba sta pokazala višjo celokupno natančnost kot strokovnjaki z vidno oceno (78 % in 80 % proti 71 %). Zaključki: Identificirali in validirali smo presnovna možganska vzorca pri AD in DLB. Pokazali smo tako njuno diagnostično vrednost in biološko pomembnost kot uporabnost v avtomatiziranem računalniško podprtem diferencialno diagnostičnem orodju. Novi avtomatizirani klasifikacijski algoritem je dosegel odlično diagnostično vrednost in nudil dodatne informacije vidni oceni.

• Language: Slovenian
• Keywords: Nevrodegenerativne demence, biološki označevalec, FDG PET, mrežna analiza, strojno učenje.
• Work type: Doctoral dissertation
• Organization: MF - Faculty of Medicine
• Year: 2023
• PID: 20.500.12556/RUL-145580
• COBISS.SI-ID: 150110211
• Publication date in RUL: 23.04.2023

Secondary language

• Language: English
• Title: Identification of specific metabolic brain patterns of common dementias and their application to computer-aided differential diagnostic tools

Abstract

Background: Differential diagnosis of neurodegenerative dementias may be challenging in the earliest disease stages; therefore, the use of biomarkers is becoming common in clinical practice and research. Improvement of current and development of new biomarkers are priorities in neuroscience and clinical neurology to improve diagnostic accuracy. Functional brain imaging with 2-[18F]fluoro-2-deoxy-D-glucose and positron emission tomography (2-[18F]FDG PET) can detect alterations in brain metabolism, reflecting the underlying neurodegeneration. By the application of network analysis on the 2-[18F]FDG PET images, specific and disease-related metabolic brain patterns can be identified. These patterns can be quantified on a single patient basis and resulting scores can be used for differential diagnosis, prognosis, tracking disease progression and, eventually, response to disease-modifying treatment. Aims: to identify and validate specific metabolic brain patterns characteristic for the two most common neurodegenerative dementias: dementia due to Alzheimer’s disease (AD) and dementia with Lewy bodies (DLB) by applying network analysis to 2-[18F]FDG PET brain scans. Additionally, we aimed to develop an automated classification algorithm for the differentiation and classification of 2-[18F]FDG PET images from various dementia patients and to compare its diagnostic performance to experts’ visual reading. Methods: We analysed clinical data and 2-[18F]FDG PET brain scans of 388 subjects: 133 patients with AD, 42 patients with mild cognitive impairment (MCI), 79 patients with DLB, 23 patients with behavioural variant frontotemporal dementia (bvFTD) and 111 normal controls (NCs). After image pre-processing, (1.) scaled subprofile model/principal component analysis (SSM/PCA) was applied to 20 AD and 20 NC 2-[18F]FDG PET scans to identify AD-related pattern (ADRP). Pattern’s diagnostic accuracy was calculated and validated on three independent validation cohorts. Further, the pattern’s expression scores were correlated with patients’ global cognitive measures and disease duration; (2.) SSM/PCA was applied to 20 DLB and 20 NC 2-[18F]FDG PET brain scans to identify DLB-related pattern (DLBRP). Pattern’s diagnostic accuracy was calculated and validated on an independent validation cohort of 59 DLB, 63 AD and 21 NC scans and pattern’s expression scores were further correlated with patients’ global cognitive measures and disease duration; (3.) a machine learning model employing two sets of features (an expression of disease-related patterns and mean regional glucose uptake values in 95 regions of interest (ROIs)) was built and tested. The diagnostic performance of classification algorithms was compared to the experts’ visual readings. A clinical diagnosis at follow-up (M = 25 ± 20 months) or cerebrospinal fluid biomarkers for Alzheimer’s disease were considered a diagnostic gold standard. Results: (1.) The newly derived ADRP was characterized by a relatively reduced metabolic activity in temporoparietal cortices, posterior cingulate and precuneus which co-varied with relatively increased activity in the cerebellum. ADRP expression significantly differentiated AD from NC (areas under the curves (AUCs) = 0.95–0.98) and correlated with measures of global cognition indices in all studied AD patient cohorts; (2.) The newly derived DLBRP was characterized by relatively reduced metabolic activity in occipital, inferior parietal and inferior temporal cortices and precuneus which co-varied with relatively increased metabolic activity in basal ganglia and mesial temporal cortex. DLBRP expression significantly differentiated DLB from NC (AUC = 0.99) and correlated with patients’ measures of global cognition. After accounting for the topographic overlap of ADRP with DLBRP, also DLB could be significantly differentiated from AD (AUC = 0.87); (3.) The pattern- and ROI-based classifier achieved a high diagnostic accuracy with high specificities and sensitivities in all four participant groups (AD, DLB, bvFTD and NC). Both classifiers achieved a higher overall diagnostic accuracy than expert readers (78% and 80% respectively, vs. 71%). Conclusions: We identified and validated metabolic brain biomarkers of AD and DLB. We showed their diagnostic value and biological significance as well as their utility in an automated computer-aided differential diagnostic tool. A newly built automated classification algorithm has shown an excellent diagnostic performance and it provided additional information to visual reading.

• Keywords: Neurodegenerative dementias, biomarker, FDG PET, network analysis, machine learning.