Projects

As part of the BiRT project, we have acquired a comprehensive imaging dataset from 70 prostate cancer patients including multiparametric MRI, ex vivo MRI, PET/CT and ground truth histology data. Using our sophisticated co-registration framework, we are conducting advanced correlation analyses using radiomics methods to validate mpMRI and PET with histology. These advanced analyses will provide important supportive information for BiRT treatment planning.

Machine learning techniques have been used to estimate the tumour location as well as its biological characteristics from the multiparametric MRI data. The intention is to extract patient-specific information from medical imaging which can be used to personalise the radiation therapy treatment. This approach applies a non-uniform dose distribution to the prostate, which can spare healthy tissues while maintaining tumour control.

Our personalised approach with a non-uniform dose distribution has been compared against the conventional treatment planning. The result shows an improved tumour control probability (TCP) while sparing the healthy tissues. The BiRT principles were applied to prostate intensity-modulated radiotherapy (IMRT) to assess the feasibility and dosimetric benefits of the proposed approach compared to uniform-dose planning. The bio-focused plans demonstrated potential in producing dose distributions with high tumour control and lower rectal and bladder toxicity compared to dose-based planning methods, across several dose fractionation schedules. This is an advance compared with the current non-discriminative approach which delivers a uniform high dose to the entire prostate.

Patient follow-up data has been analysed to track the changes on the multiparametric MRI data. The aim is to identify imaging biomarkers which can predict treatment outcome and patient survival. Imaging biomarkers are characteristic features in the image which can be predictive or prognostic. In this ongoing project, a panel of radiomics features have been computed from the imaging data, which are currently studied over the course of follow-up.

We have designed and performed studies to bridge the gap between the imaging phenotype and the underlying biological genotype. Punch biopsies were taken from the FFPE samples which were sent for transcriptome sequencing. Meanwhile, the corresponding MRI data was analysed by extracting texture features. Weak correlations were found between the imaging features and certain gene expressions. To reduce the false discovery due to the high dimensionality of the genetic data, this study focuses on the hypoxia-related gene set. This ties with the general interest in hypoxia of prostate cancer, which is still currently under-studied.

Nym Venderberg’s work focuses on the use of machine learning and deep learning to predict and grade prostate cancer in histopathology data, in order to build models that will be incorporated into the BiRT framework. They have previously explored the use of support vector machines, on which they presented at EPSM 2018, and they are currently investigating applications of transfer learning methods to refine their histopathology prediction models.

Whilst most radiotherapy treatments are delivered using high energy, X-ray producing machines (called linear accelerators, or linacs for short), radiation can also be applied by implanting radioactive sources into the tumour, known as brachytherapy. These radioactive sources may either be implanted permanently or removed after the radiation has been delivered. Because the radioactive source is implanted directly in the tumour, the healthy tissue surrounding the tumour receives very little radiation. The BiRT team have developed dose optimisation methods that determine precisely where the radioactive sources should be implanted in the tumour to deliver the optimal dose of radiation. (Image courtesy: John Betts)