Learning-based tumor segmentation using metabolic imaging features

The ultimate objective of brain tumor imaging is to distill patient-specific knowledge that guides therapy planning and medical care. Magnetic resonance imaging (MRI) is a prevailing technique to visualize tumors due to its excellent contrast of soft tissue and non-invasiveness. Decades of research have helped brain tumor segmentation performance dramatically. However, precise segmentation is still considered hard partly due to the limitation in resolution, signal-to-noise ratio, and possible artifacts. While some tumors are easier to delineate, more infiltrating ones like gliomas have ragged and obscure boundaries that are harder to define. In recognition of this hardship, researchers have started exploring the use of Proton Magnetic Resonance Spectroscopic Imaging (MRSI) for better tumor prognosis, diagnosis, and characterization. MRSI investigates the spatial distribution of metabolic changes by leveraging its unique temporal information. The wealth of this spectroscopic information is beneficial in classifying tumor subregions and aiding ongoing research investigations in tumor heterogeneity and related topics. Several studies have reported an increase in choline-containing compounds level and a reduced N-acetyl-aspartate level in brain tumors. Spectroscopic techniques can pick up these metabolic changes, and they might be the missing pieces of better MRI-based tumor segmentation solutions. This study shows a successful application of deep learning and MRSI to identify tumor and edema regions of human brains with glioblastomas. The deep learning framework of choice is nnU-Net. Most specialized solutions in applying deep neural models in the medical image domain depend on dataset properties and hardware constraints. nnU-Net is a framework that automatically adapts itself to various medical image segmentation tasks. Therefore, it ensures a fair comparison of experiments. This work shows an improved segmentation result after incorporating high-resolution metabolite maps derived from MRSI data acquired by the SPICE sequence. The high resolution, rapidness, and near whole-brain performance of SPICE should assist radiologists and oncologists in delimiting the pathological area better and providing more appropriate medical help.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Liang, Yucheng

Contributors dc:contributor

• Liang, Zhi-Pei

• Engineering

Statement dc:rights

• Copyright 2021 Yucheng Liang

Language dc:language

en, eng

Handle dc:identifier

http://hdl.handle.net/2142/113936

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F113936