Sir Peter MacCallum Department of Oncology - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemDetection and characterisation of prostate cancer from multiparametric MRI using machine learning techniques( 2018)Prostate cancer (PCa) is the most commonly diagnosed cancer type in males in Australia and radiotherapy provides an efficacious treatment option. However, traditional approaches of radiotherapy apply a uniform high radiation dose to the entire prostate, which can cause side effects as healthy tissues are irradiated. The treatment-related side effects have a substantial impact on patients’ quality of life, particularly for the increasing number of low-risk and intermediate-risk patients who typically have a long life-expectancy (10 – 20 years). Focal therapy provides a solution to reduce treatment-related side effects by selectively treating the tumour subvolume and hence spares healthy tissues. Our group has proposed a specific form of focal therapy, termed “bio-focused radiotherapy” (BiRT), where the actual radiation dose to be delivered to the tumour volume further depends on the biological characteristics of the tumour, including cell density, tumour aggressiveness and the presence of hypoxia. To implement BiRT, reliable estimation of tumour location and characteristics is required. The aim of this thesis is to develop predictive models for tumour location and biological characteristics from multiparametric MRI (mpMRI) for the clinical implementation of BiRT. In vivo mpMRI scans were acquired from 30 patients who subsequently underwent radical prostatectomy. Histology was obtained from prostate specimens after surgery, which was used to retrieve the “ground truth” information of tumour location and characteristics for predictive model development using supervised learning. To enable voxel-wise analysis, in vivo mpMRI data was co-registered with ex vivo histology using an advanced 3D deformable registration framework. This provided the necessary materials for predictive model development, using mpMRI data as features (predictors) and ground truths from histology as labels. To facilitate the investigation of specific models, a multi-purposed scalable machine learning workflow (SMLW) was first developed. Preliminary investigations were performed to examine the usefulness of the SMLW in predicting tumour location at a voxel level. After validation of the methodology, following investigations focused on the prediction of prostate cell density and tumour aggressiveness. To address the challenges associated with traditional approaches for measuring hypoxia, radiogenomics approaches were applied to enable the integrative analysis of the mpMRI and the corresponding genomic profiles assessed using next generation sequencing. Investigation into predicting tumour location using the SMLW achieved a model with an area under the receiver operating characteristics curve (AUC of ROC) ranging from 0.81 to 0.94, which was comparable with previous literature. A predictive model for cell density prediction was also developed and gave a root mean square error of 1.06 x 10^3 cells / mm^2, equivalent to a relative error of 13.25%. Results for predicting tumour aggressiveness achieved an AUC of 0.91, and two high-performance run length and size zone texture features were identified as promising biomarkers. Radiogenomics analysis of prostate hypoxia revealed a selection of 16 texture features which showed weak but significant correlations with hypoxia-related gene expression levels. Overall, this thesis presents the development of a flexible machine learning workflow, which was applied with specific settings to predict tumour location and characteristics. The goal was to enable non-invasive detection and characterisation of PCa for personalised radiotherapy treatment optimization using the BiRT approach. In addition, this thesis also aimed to fill gaps within the current prostate radiomics community. This included the development of the first model for prostate cell density, a systematic comparison of texture features for predicting tumour aggressiveness and the first study on prostate hypoxia using radiogenomics approaches. The contributions from this thesis may facilitate the continuous development of methods for the prostate radiomics community.