Advanced image analysis to track daily biological changes in radiotherapy.

Background: Radiotherapy in the treatment of cancer using high-energy x-rays. For patients with prostate cancer, radiotherapy is one of the main treatment options, used in combination with hormonal treatments. Treatment outcomes are good for prostate cancer with many men cured. It is likely that for some men we are over treating their disease and for other men we are under treating their disease which results in treatment failure.

Radiotherapy is delivered to a patient across a number of fractions, treated over a number of days. The two main fractionation schedules are to treat over 20 days or a high dose given in a shorter time over 5 days. On each treatment fraction an image of the patient is taken to ensure they are aligned correctly for treatment. This image guidance may be a cone beam CT or now, may be MR guided. These images may contain additional information about how the prostate cancer is responding to treatment which can be used to adapt the treatment (treatment escalation or de-escalation) for men on treatment.

Hypothesis: MR images taken during treatment contain information on how the prostate cancer is responding to treatment which can be used to personalize treatment for every patient.

This project will develop the image analysis pipeline required to perform the analysis longitudinal medical image data. This will use our in-house research platform which includes image registration (rigid and non-rigid), contour propagation and image processing tools and will be linked to both (1) pyradiomics, an open-source python package, which extracts common pre-defined image features and (2) machine learning approaches.

Objectives: The following objectives will be addressed in this project (each is a planned paper):
1. Develop a methodology to handle longitudinal imaging data, selecting the image features that best describe the tumour environment. Image features in time-series data will be modelled with Gaussian process to handle any irregularities in the time-series. Hierarchical clustering will group features with similar trajectories and cross-correlations will select the best, representative, feature for each cluster.
2. Image normalisation is needed to remove any acquisition related offsets in the imaging data. This ensures that image features represent true changes in the tissue and not day-to-day machine acquisition changes. Several normalisation approaches will be tested to define a best practice approach.
3. Using the above defined methodologies we will first investigate longitudinal changes in the prostate cancer versus normal prostate tissue. A clinical expert has delineated the dominate prostate cancer region allowing image features to be extracted.
4. Investigate correlations between known biological states of the prostate (i.e., if the prostate is hypoxic, low oxygenation, defined from a genomic signature calculated form the cancer biopsy).

The proposed project aligns with the healthcare technology's theme, and then with two strategies within this theme:
Frontiers of Physical Intervention + Optimising interventions:
By linking the patients imaging phenotype with underlying biology of their disease we can minimise the need
for invasive sample collection, maximise the number of time-points available and speed the availability of data
for optimising treatment interventions.