Myopia, whose prevalence is rising worldwide, is characterised by axial elongation of the globe that remodels the posterior pole; among the affected tissues, the choroid is the first to thin, to the point of being regarded as a biomarker of myopic progression. Quantifying it on OCT images — choroidal thickness (CT), subfoveal thickness (SFChT), area and choroidal vascularity index (CVI) — is hindered by manual segmentation, which is labour-intensive and subject to inter-operator variability. Deep-learning solutions for semantic segmentation of the choroid, such as Choroidalyzer (a three-channel U-Net: region, vessel, fovea), are trained on general populations and offer no performance guarantee on the highly myopic eye, where the choroid is thin and the chorioscleral junction is poorly contrasted (domain shift). This thesis addresses the problem with a two-stage domain-adaptation pipeline based on transfer learning: starting from the official pre-trained weights, an initial (baseline) training on the general domain is followed by fine-tuning on the myopic domain. The ground truth is built by generating pseudo-masks with the official weights and manually correcting them (gold standard); the train/validation/test split is frozen and reproducible. The adapted model is evaluated with Dice and IoU against the official weights and the baseline, with automatic extraction of the biomarkers and assessment of their agreement with manual measurements. The work is rounded off by a proof-of-concept longitudinal analysis with a foveal lock, which anchors the measurements to the same foveal coordinate to estimate percentage changes over time.
La miopia, in rapida crescita a livello mondiale, è caratterizzata da un allungamento assiale del bulbo che rimodella il polo posteriore; tra i tessuti coinvolti, la coroide è quello che si assottiglia più precocemente, tanto da essere considerata un biomarcatore della progressione miopica. La sua quantificazione su immagini OCT — spessore coroideale (CT), spessore subfoveale (SFChT), area e indice di vascolarizzazione (CVI) — è però ostacolata dalla segmentazione manuale, onerosa e soggetta a variabilità inter-operatore. Le soluzioni di deep learning per la segmentazione semantica della coroide, come Choroidalyzer (una U-Net a tre canali: region, vessel, fovea), sono addestrate su popolazioni generali e non offrono garanzie di prestazione sull'occhio fortemente miope, dove la coroide è sottile e la giunzione corio-sclerale è poco contrastata (domain shift). Questa tesi affronta il problema con una pipeline di domain adaptation a due stadi basata sul transfer learning: a partire dai pesi ufficiali pre-addestrati, un primo addestramento (baseline) sul dominio generale è seguito da un fine-tuning sul dominio miopico. Il ground truth è costruito generando pseudo-maschere con i pesi ufficiali e correggendole manualmente (gold standard); la suddivisione train/validation/test è congelata e riproducibile. Il modello adattato è valutato in Dice e IoU contro i pesi ufficiali e la baseline, con estrazione automatica dei biomarcatori e verifica dell'accordo con le misure manuali. Completa il lavoro un proof of concept di analisi longitudinale con «foveal lock», che àncora le misure alla stessa coordinata foveale per stimare le variazioni percentuali nel tempo.
MACHINE LEARNING PER LA SEGMENTAZIONE AUTOMATICA DELLA COROIDE IN IMMAGINI OCT DI SOGGETTI MIOPI
SERRA, FEDERICA
2025/2026
Abstract
Myopia, whose prevalence is rising worldwide, is characterised by axial elongation of the globe that remodels the posterior pole; among the affected tissues, the choroid is the first to thin, to the point of being regarded as a biomarker of myopic progression. Quantifying it on OCT images — choroidal thickness (CT), subfoveal thickness (SFChT), area and choroidal vascularity index (CVI) — is hindered by manual segmentation, which is labour-intensive and subject to inter-operator variability. Deep-learning solutions for semantic segmentation of the choroid, such as Choroidalyzer (a three-channel U-Net: region, vessel, fovea), are trained on general populations and offer no performance guarantee on the highly myopic eye, where the choroid is thin and the chorioscleral junction is poorly contrasted (domain shift). This thesis addresses the problem with a two-stage domain-adaptation pipeline based on transfer learning: starting from the official pre-trained weights, an initial (baseline) training on the general domain is followed by fine-tuning on the myopic domain. The ground truth is built by generating pseudo-masks with the official weights and manually correcting them (gold standard); the train/validation/test split is frozen and reproducible. The adapted model is evaluated with Dice and IoU against the official weights and the baseline, with automatic extraction of the biomarkers and assessment of their agreement with manual measurements. The work is rounded off by a proof-of-concept longitudinal analysis with a foveal lock, which anchors the measurements to the same foveal coordinate to estimate percentage changes over time.| File | Dimensione | Formato | |
|---|---|---|---|
|
TESI_SERRA.pdf
File consultabile da postazioni biblioteca
Descrizione: Tesi definitiva, in pdf.
Dimensione
5.66 MB
Formato
Adobe PDF
|
5.66 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/111024