| dc.description.abstract | The primary function of a predictive biomarker is to accurately determine the outcome of a specified treatment. Radiomics is a developing field that involves advanced image analysis and high throughput extraction of mineable precise quantitative imaging descriptors or features that serve as non-invasive prognostic or predictive biomarkers.
Difficulty in attaining reproducibility of radiomics features has hindered their widespread use in the clinic to date. In Study 1 of this thesis, the impact of image acquisition parameters on feature reproducibility in T2-weighted MRI is analysed using a custom-designed 3D printed prostate phantom. This study concludes that the parameter with the greatest influence on feature reproducibility is slice thickness and that when a reasonable level of spatial resolution has been achieved, matrix size and field of view have limited impact on feature reproducibility.
In Study 2, the impact of image pre-processing on reproducibility of radiomics features in multiparametric MRI for prostate cancer is explored. From this work, it can be concluded that image normalisation and voxel resampling are critical to reproducible feature extraction and that linear interpolation algorithm appears favourable to nearest neighbour, particularly in the reproducibility of textural features.
Radiation therapy, using a variety of delivery methods, is a definitive management modality in treating prostate cancer. One of the main challenges in the management of prostate cancer is the stratification of those with clinically significant disease from those with indolent disease, as the treatment trajectories of such patients differ. Another issue is the potential to identify and dose escalate the hypoxic sub-volume, a known cause of radiation therapy treatment failure. Studies 3 and 4 of this thesis explore if radiomics features extracted from multiparametric MRI images can predict for pimonidazole (PIMO) scoring as a surrogate for hypoxia as well as some of the current clinical endpoints used in the stratification of prostate cancer (Gleason score at biopsy, clinical T stage and D amico risk stratification score). Study 3, based on T2-weighted MRI, reports statistically significant models for prediction of PIMO score, Gleason score at biopsy, clinical T stage and D amico risk classification score. ROC analysis of these models identifies that their predictive potentials range from equivocal for clinical T stage (AUC=0.501) to moderate for D amico risk classification score (AUC=0.753).
Study 4, based on DW-MRI and ADC maps, reports that the first order feature Stats_Var and the second order feature GLCM_InverseVar are present in all models identified predicting for PIMO score, Gleason score at biopsy, clinical T stage and D amico risk stratification score. This indicates that they may have potential in the stratification of clinically significant from indolent prostate cancer. The textural feature GLRLM_RLV is identified in 4 models predicting for PIMO score, with the DW-MRI radiomics-only model yielding the highest predictive potential, with an AUC of 0.582. D amico risk stratification score is predicted by models that included the morphological feature Shape_AreaDensityBB, with AUCs of 0.519 and 0.638 for DW-MRI and ADC maps, respectively.
This thesis reports reproducibility considerations for radiomics features in multiparametric MRI for prostate cancer. It also reports models with low to moderate prediction for PIMO score in prostate cancer, as well as clinical endpoints currently used in the stratification of prostate cancer patients. These models require external validation. | en |
