Neuropsychological tests such as the Trail Making Test part A (TMT-A) and the Rey Auditory Verbal Learning Test (RAVLT) are the clinical reference for quantifying cog- nitive performance, but they are slow and require trained personnel. The retina shares its embryological origin with the brain and is imaged non-invasively with Optical Co- herence Tomography (OCT) and OCT Angiography (OCTA), and a growing oculomics literature links retinal biomarkers to neurodegeneration. This thesis, carried out within the DINEURET project, asks whether OCT and OCTA images encode measurable infor- mation about cognitive performance. The question is cast as regression of continuous TMT-A and RAVLT scores rather than disease classification. The central difficulty is to separate any retinal signal from what age, education and acquisition centre already explain, on a small multicentre cohort (N ≈128 training subjects, three centres) whose subject-level labels are attached to eye-level scans. The approach rests on three formulations evaluated under one discipline: image-only ˆ y= f (x), covariate-onlyˆ y= g(c) and combinedˆ y= h(x, c), with a group-aware split that keeps all eyes, views and subjects of a centre in one partition. Several image rep- resentations, learned and interpretable, guard against a model-specific null result, and confounding is probed with demographic residualisation, domain-adversarial training and leave-one-centre-out (LOCO) evaluation. Retinal images carry a modest signal, mainly for memory-related RAVLT-immediate (R2 ≈0.22), weaker than demographics alone (R2 ≈0.46), while TMT-A is essentially unpredictable; part of the apparent association is tied to acquisition centre and collapses under LOCO. Fusing retinal representations with demographics reaches the best result (R2 ≈0.57), indicating non-redundant information. The contribution is a feasibility study and a benchmark, not a clinical screening tool.
I test neuropsicologici come il Trail Making Test parte A (TMT-A) e il Rey Auditory Verbal Learning Test (RAVLT) sono il riferimento clinico per quantificare le prestazioni cognitive, ma richiedono tempo e personale addestrato. La retina condivide l’origine embriologica con il cervello e si acquisisce in modo non invasivo con l’Optical Coherence Tomography (OCT) e l’OCT Angiography (OCTA), e una crescente letteratura di oculom- ica collega i biomarcatori retinici alla neurodegenerazione. Questa tesi, svolta nell’ambito del progetto DINEURET, si chiede se le immagini OCT e OCTA codifichino informazioni misurabili sulle prestazioni cognitive. Il problema è formulato come regressione dei punteggi continui TMT-A e RAVLT, non come classificazione di malattia. La difficoltà centrale è separare l’eventuale segnale re- tinico da ciò che età, istruzione e centro di acquisizione già spiegano, su una coorte multicentrica piccola (N ≈128 soggetti di training, tre centri) in cui le etichette a livello di soggetto sono associate a scansioni a livello di occhio. L’approccio si fonda su tre formulazioni valutate con la stessa disciplina: solo-immagine ˆ y= f (x), solo-covariateˆ y= g(c) e combinataˆ y= h(x, c), con uno split group-aware per occhi, viste e soggetti di un centro. Diverse rappresentazioni, apprese e interpretabili, evitano che un risultato nullo dipenda da una sola famiglia di modelli, e il confondi- mento è indagato con residualizzazione demografica, addestramento domain-adversarial e valutazione leave-one-centre-out (LOCO). Le immagini retiniche portano un segnale modesto, principalmente per la memoria mis- urata da RAVLT-immediate (R2 ≈0.22), più debole delle sole covariate demografiche (R2 ≈0.46), mentre il TMT-A è di fatto non predicibile; parte dell’associazione apparente è legata al centro di acquisizione e crolla in regime LOCO. La fusione delle rappresentazioni retiniche con le covariate demografiche raggiunge il risultato migliore (R2 ≈0.57), segno di informazione non ridondante. Il contributo è uno studio di fattibilità e un benchmark, non uno strumento di screening clinico.
DINEURET: does the retina encode cognition? An OCT/CTA feasibility study
Ferraù, Francesco
2025/2026
Abstract
Neuropsychological tests such as the Trail Making Test part A (TMT-A) and the Rey Auditory Verbal Learning Test (RAVLT) are the clinical reference for quantifying cog- nitive performance, but they are slow and require trained personnel. The retina shares its embryological origin with the brain and is imaged non-invasively with Optical Co- herence Tomography (OCT) and OCT Angiography (OCTA), and a growing oculomics literature links retinal biomarkers to neurodegeneration. This thesis, carried out within the DINEURET project, asks whether OCT and OCTA images encode measurable infor- mation about cognitive performance. The question is cast as regression of continuous TMT-A and RAVLT scores rather than disease classification. The central difficulty is to separate any retinal signal from what age, education and acquisition centre already explain, on a small multicentre cohort (N ≈128 training subjects, three centres) whose subject-level labels are attached to eye-level scans. The approach rests on three formulations evaluated under one discipline: image-only ˆ y= f (x), covariate-onlyˆ y= g(c) and combinedˆ y= h(x, c), with a group-aware split that keeps all eyes, views and subjects of a centre in one partition. Several image rep- resentations, learned and interpretable, guard against a model-specific null result, and confounding is probed with demographic residualisation, domain-adversarial training and leave-one-centre-out (LOCO) evaluation. Retinal images carry a modest signal, mainly for memory-related RAVLT-immediate (R2 ≈0.22), weaker than demographics alone (R2 ≈0.46), while TMT-A is essentially unpredictable; part of the apparent association is tied to acquisition centre and collapses under LOCO. Fusing retinal representations with demographics reaches the best result (R2 ≈0.57), indicating non-redundant information. The contribution is a feasibility study and a benchmark, not a clinical screening tool.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Ferraù.pdf
non accessibile
Descrizione: Testo della Tesi
Dimensione
30.47 MB
Formato
Adobe PDF
|
30.47 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Ferraù_Executive Summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Testo dell'Executive Summary
Dimensione
2.39 MB
Formato
Adobe PDF
|
2.39 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/10589/260798