Non-small cell lung cancer computed tomography imaging: development of analytical tools to improve segmentation and prognosis
File(s)
Author(s)
Boubnovski Martell, Marc
Type
Thesis or dissertation
Abstract
Advanced diagnostic and prognostic tools are imperative for Non-Small Cell Lung Cancer (NSCLC) given its high prevalence and diagnostic complexity. This research introduces novel
image analysis methodologies for NSCLC, aiming to enhance Computed Tomography (CT) image interpretability by focusing on the biophysical texture properties within lung, key among
which are automatic lobar segmentation, image quality enhancement, and nodule feature quality assessment. Models were rigorously tested using external datasets, including over 500 patients
from three centres, to ensure robustness. Development of a V-Net model, culminated in lung lobar segmentation across normal, COPD, lung cancer, and COVID-19 pneumonitis conditions, achieving high Dice scores (up to 0.97), significant at p < 0.05. This model is expected to further aid current nodule segmentation efforts. Regarding nodule characterisation, a Deep Probabilistic Canonical Correlation Analysis (DPCCA) trained on CT and metabolomics data extracted Tissue Metabolomic Radiomics-CT (TMR-CT) features, correlating with metabolic profiles (correlation coecients up to 0.61). TMR-CT significantly outperformed traditional radiomics in NSCLC prognosis, achieving an F1-score of 0.78 for histological classification and a prognosis c-index of 0.72. To verify if these improvements will benefit from resolution increases and harmonisation, a Transformer Volumetric Super-Resolution Network (TVSRN-V2) was developed, which improved image structural similarity index (SSIM) of 0.941±0.028 and peak signal-to-noise ratio (PSNR) of 39.162±1.843, enhancing small region segmentation and prognosis accuracy. This research merges advanced imaging and machine learning for NSCLC texture analysis, refining lobar segmentation, and improving tumour localisation and prognosis predictions. The comprehensive use of CT datasets and integration of high-resolution imaging analytics deepens the understanding of lung NSCLC texture, potentially advancing CT-based prognosis and treatment planning tools.
image analysis methodologies for NSCLC, aiming to enhance Computed Tomography (CT) image interpretability by focusing on the biophysical texture properties within lung, key among
which are automatic lobar segmentation, image quality enhancement, and nodule feature quality assessment. Models were rigorously tested using external datasets, including over 500 patients
from three centres, to ensure robustness. Development of a V-Net model, culminated in lung lobar segmentation across normal, COPD, lung cancer, and COVID-19 pneumonitis conditions, achieving high Dice scores (up to 0.97), significant at p < 0.05. This model is expected to further aid current nodule segmentation efforts. Regarding nodule characterisation, a Deep Probabilistic Canonical Correlation Analysis (DPCCA) trained on CT and metabolomics data extracted Tissue Metabolomic Radiomics-CT (TMR-CT) features, correlating with metabolic profiles (correlation coecients up to 0.61). TMR-CT significantly outperformed traditional radiomics in NSCLC prognosis, achieving an F1-score of 0.78 for histological classification and a prognosis c-index of 0.72. To verify if these improvements will benefit from resolution increases and harmonisation, a Transformer Volumetric Super-Resolution Network (TVSRN-V2) was developed, which improved image structural similarity index (SSIM) of 0.941±0.028 and peak signal-to-noise ratio (PSNR) of 39.162±1.843, enhancing small region segmentation and prognosis accuracy. This research merges advanced imaging and machine learning for NSCLC texture analysis, refining lobar segmentation, and improving tumour localisation and prognosis predictions. The comprehensive use of CT datasets and integration of high-resolution imaging analytics deepens the understanding of lung NSCLC texture, potentially advancing CT-based prognosis and treatment planning tools.
Version
Open Access
Date Issued
2024-04-08
Date Awarded
2024-08-01
Copyright Statement
Creative Commons Attribution Licence
License URL
Advisor
Aboagye, Eric
Sponsor
Medical Research Council (Great Britain)
Publisher Department
Department of Surgery & Cancer
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)