This thesis explores the use of three-dimensional convolutional neural networks for the automatic segmentation of brain tumors from magnetic resonance imaging (MRI). The main goal is to evaluate how di erent strategies for data representation and manipulation can enhance the ability of neural networks to accurately identify and delineate tumor regions. In order to achieve this, based on the public dataset from the FeTS Challenge 2022, two complementary approaches were investigated: the use of Signed Distance Maps, which provide a continuous description of the transition between healthy and pathological tissue, and SynthStrip-like augmentations, designed to generate synthetic images and increase data variability. The analysis involved UNet-based architectures, specifically the 3D UNet and TokenUNet, whose performance was evaluated through standard quantitative metrics and a representation learning study focused on feature maps, internal correlations, and principal component analysis (PCA). The results show that the use of continuous representations and synthetic data improves segmentation quality and makes the model more adaptable to real-world clinical scenarios.
Questo lavoro di tesi esplora l’utilizzo di convolutional neural networks 3D per la segmentazione automatica di tumori cerebrali da immagini di risonanza magnetica. L’obiettivo principale è valutare come diverse strategie di rappresentazione e manipolazione dei dati possano migliorare la capacità delle neural networks di riconoscere e delimitare con precisione le regioni tumorali. A tal fine, sulla base del dataset pubblico della FeTS Challenge 2022, sono state sperimentate due soluzioni complementari: l’impiego delle Signed Distance Maps, che descrivono in modo continuo la transizione tra tessuto sano e patologico, e l’uso di SynthStrip-like augmentations, volte a generare immagini sintetiche e aumentare la varietà dei dati disponibili. L’analisi ha coinvolto architetture basate su UNet, in particolare la 3D UNet e la TokenUNet, le cui prestazioni sono state analizzate con metriche quantitative standard e tramite uno studio del representation learning basato su feature map, correlazioni interne e PCA. I risultati ottenuti mostrano che l’uso di rappresentazioni continue e di dati sintetici contribuisce a migliorare la qualità delle segmentazioni e a rendere il modello più adattabile a scenari clinici reali.
Apprendimento di Rappresentazioni generali ed efficaci con Reti Neurali di tipo UNet, tramite Signed Distance Maps e Synthstrip-like augmentations
MUSU, MARIA
2024/2025
Abstract
This thesis explores the use of three-dimensional convolutional neural networks for the automatic segmentation of brain tumors from magnetic resonance imaging (MRI). The main goal is to evaluate how di erent strategies for data representation and manipulation can enhance the ability of neural networks to accurately identify and delineate tumor regions. In order to achieve this, based on the public dataset from the FeTS Challenge 2022, two complementary approaches were investigated: the use of Signed Distance Maps, which provide a continuous description of the transition between healthy and pathological tissue, and SynthStrip-like augmentations, designed to generate synthetic images and increase data variability. The analysis involved UNet-based architectures, specifically the 3D UNet and TokenUNet, whose performance was evaluated through standard quantitative metrics and a representation learning study focused on feature maps, internal correlations, and principal component analysis (PCA). The results show that the use of continuous representations and synthetic data improves segmentation quality and makes the model more adaptable to real-world clinical scenarios.| File | Dimensione | Formato | |
|---|---|---|---|
|
Musu_Maria.pdf
accesso aperto
Dimensione
9.53 MB
Formato
Adobe PDF
|
9.53 MB | Adobe PDF | Visualizza/Apri |
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/20.500.12608/98052