Geographic atrophy (GA) is an advanced form of age-related macular degeneration (AMD) and is a leading cause of visual impairment in the elderly. Precise segmentation of GA lesions is fundamental for monitoring disease progression in both clinical trials and daily ophthalmological practice. However, manual delineation is a complex, time-consuming process and is subject to inter-observer variability. In recent years, artificial intelligence (AI) has shown great potential in the medical field, offering tools capable of improving efficiency and reducing the clinical workload. [4] The goal of this thesis is to develop an algorithm using AI for the automatic segmentation of GA lesions, analyzing its performance through metrics such as the Dice similarity coefficient. This thesis presents the development of different methodologies to address the disease segmentation problem, comparing their advantages and critical aspects in order to identify the most suitable approach. The aim is to contribute to the advancement of automated solutions that can be truly implemented in clinical practice, bridging the existing gap between research and real-world application. The proposed models were trained, validated, and tested using proprietary datasets provided by CENTERVUE S.P.A. and manually segmented by a retina specialist. The data includes images acquired with the Eidon and DRSplus devices, both TrueColor confocal fundus imaging systems designed and developed by the company. The results show that a net based on the U-Net architecture, when combined with postprocessing, can be considered an appropriate method to perform GA segmentation from fundus images.
L’Atrofia Geografica (GA) e una forma avanzata della degenerazione maculare legata all’età (AMD) ed e una causa principale di compromissione visiva negli anziani. La segmentazione precisa delle lesioni di GA e fondamentale per il monitoraggio della progressione della malattia sia negli studi clinici che nella pratica oftalmologica quotidiana. Tuttavia, la delineazione manuale e un processo complesso, che richiede tempo, ed è soggetta a variabilità tra osservatori. Negli anni recenti, l’intelligenza artificiale (AI) ha mostrato grande potenziale nel campo medico, offrendo strumenti capaci di migliorare l’efficienza e di ridurre il carico di lavoro clinico. [4] L’obiettivo di questa tesi e sviluppare un algoritmo che utilizzi l’IA per la segmentazione automatica delle lesioni di GA, analizzando la sua performance attraverso metriche come il coefficiente di similarita di Dice. Questa tesi presenta lo sviluppo di diverse metodologie per affrontare il problema della segmentazione della malattia, confrontando i loro vantaggi e gli aspetti critici al fine di identificare l’approccio piu adatto. Lo scopo e contribuire all’avanzamento di soluzioni automatizzate che possano essere veramente implementate nella pratica clinica, colmando il divario esistente tra la ricerca e l’applicazione nel mondo reale. I modelli proposti sono stati addestrati, validati e testati usando dataset proprietari forniti da CENTERVUE S.P.A. e segmentati manualmente da un retinologo specialista. I dati includono immagini acquisite dai dispositivi Eidon e DRSplus, entrambi sistemi di imaging confocale del fondo oculare TrueColor progettati e sviluppati dell’azienda. I risultati dimostrano che una rete basata sull’architettura U-Net, se combinata con il postprocessing, puo essere considerata un metodo appropriato per eseguire la segmentazione della GA da immagini del fundus.
Segmentazione AI-based di atrofia geografica su immagini del fundus della retina. Modelli di intelligenza artificiale su un dataset di immagini retiniche affette da atrofia geografica segmentato da specialista della retina
MENOTTI, REBECCA
2024/2025
Abstract
Geographic atrophy (GA) is an advanced form of age-related macular degeneration (AMD) and is a leading cause of visual impairment in the elderly. Precise segmentation of GA lesions is fundamental for monitoring disease progression in both clinical trials and daily ophthalmological practice. However, manual delineation is a complex, time-consuming process and is subject to inter-observer variability. In recent years, artificial intelligence (AI) has shown great potential in the medical field, offering tools capable of improving efficiency and reducing the clinical workload. [4] The goal of this thesis is to develop an algorithm using AI for the automatic segmentation of GA lesions, analyzing its performance through metrics such as the Dice similarity coefficient. This thesis presents the development of different methodologies to address the disease segmentation problem, comparing their advantages and critical aspects in order to identify the most suitable approach. The aim is to contribute to the advancement of automated solutions that can be truly implemented in clinical practice, bridging the existing gap between research and real-world application. The proposed models were trained, validated, and tested using proprietary datasets provided by CENTERVUE S.P.A. and manually segmented by a retina specialist. The data includes images acquired with the Eidon and DRSplus devices, both TrueColor confocal fundus imaging systems designed and developed by the company. The results show that a net based on the U-Net architecture, when combined with postprocessing, can be considered an appropriate method to perform GA segmentation from fundus images.| File | Dimensione | Formato | |
|---|---|---|---|
|
Menotti_Rebecca.pdf
Accesso riservato
Dimensione
39.58 MB
Formato
Adobe PDF
|
39.58 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/95803