Politecnico di Torino (logo)

Deep neural networks for brain tumor segmentation in magnetic resonance imaging

Luca Francesco Rossi

Deep neural networks for brain tumor segmentation in magnetic resonance imaging.

Rel. Lia Morra, Fabrizio Lamberti, Luca Mainardi. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2022

PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (45MB) | Preview
[img] Archive (ZIP) (Documenti_allegati) - Other
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (784kB)

Gliomas account for approximately 30% of all brain and central nervous system tumors, and 80% of all malignant brain tumors. Glioblastoma multiforme (GBM) are the most commonly encountered and aggressive malignant primary tumor (15%) of the central nervous system in adults, accounting for approximately 55% of all gliomas. Nowadays, the standard of care treatment for GBM patients consists in surgical resection followed by radiotherapy and chemotherapy, leaving then the subject untreated for the succeeding four months. Deep Learning (DL) has achieved cutting edge results in several medical fields, with its applicability ranging from lesion segmentation to disease relapse prediction. Neuro-oncology, being one of those, has seen important advancements, especially in automating neuroradiology tasks such as brain tumor detection and segmentation. However, even if state-of-the-art results have been achieved by DL methods on brain tumor segmentation on pre-operative Magnetic Resonance Imaging (MRI) scans, hardly the same can be said of post-operative segmentation, where literature lacks of a more comprehensive study and the few proposed models still present strongly sub-human performances. Due to the lack of available data in clinical practice, Transfer Learning (TL) has seen a spike in popularity within the medical field, since it allows to train models in absence of a large dataset by leveraging knowledge learned from other source tasks. Still, current TL techniques in medical imaging mostly implement knowledge transfer from natural imaging, usually from model trained on the ImageNet dataset. Even if some progress is done, the knowledge transferred between the two areas can be either not sufficient to achieve promising results in the medical task or make the transfer process quite unpredictable. The aim of this study, in conjunction with the Molinette Hospital in Turin, is to make a further step in the field of automatic and semi-automatic brain tumor segmentation from neuroimaging modalities via deep learning technologies on post-operative scans. More in detail, this work arises upon the intuition that TL between pre- and post-operative brain tumor segmentation could lead to promising results by leveraging both the closeness of source and target domains, and the fact that the knowledge transfer process does not leave the medical field.

Relators: Lia Morra, Fabrizio Lamberti, Luca Mainardi
Academic year: 2022/23
Publication type: Electronic
Number of Pages: 101
Corso di laurea: Corso di laurea magistrale in Data Science And Engineering
Classe di laurea: New organization > Master science > LM-32 - COMPUTER SYSTEMS ENGINEERING
Aziende collaboratrici: UNSPECIFIED
URI: http://webthesis.biblio.polito.it/id/eprint/24696
Modify record (reserved for operators) Modify record (reserved for operators)