Multivariate methods for interpretable analysis of magnetic resonance spectroscopy data in brain tumour diagnosis

摘要: Malignant tumours of the brain represent one most difficult to treat types cancer due sensitive organ they affect. Clinical management pathology becomes even more intricate as tumour mass increases proliferation, suggesting that an early and accurate diagnosis is vital for preventing it from its normal course development. The standard clinical practise includes invasive techniques might be harmful patient, a fact has fostered intensive research towards discovery alternative non-invasive tissue measurement methods, such nuclear magnetic resonance. One variants, resonance imaging, already used in regular basis locate bound tumour; but complementary variant, spectroscopy, despite higher spatial resolution capability identify biochemical metabolites become biomarkers within delimited area, lags behind terms use, mainly interpretability. interpretation spectra corresponding thus interesting field automated methods knowledge extraction machine learning, always understanding secondary role human expert medical decision making. current thesis aims at contributing state art this domain by providing novel assistance radiology experts, focusing on complex problems delivering interpretable solutions. In respect, ensemble learning technique accurately discriminate amongst aggressive tumours, namely glioblastomas metastases, been designed; moreover, strategy increase stability biomarker identification means instance weighting provided. From different analytical perspective, tool based signal source separation, guided type-specific information developed assess existence tissues tumoural mass, quantifying their influence vicinity areas. This development led derivation probabilistic some separation techniques, which provide support uncertainty handling strategies estimation number differentiated analysed volumes. provided should assist experts through use tools tackling interpretability accuracy angles