About the Project
First Supervisor: Dr Isabel Dregely
Second Supervisor: Dr M. Jorge Cardoso
The aim of this project is to improve cancer treatment and cure rates through highly individualised radiotherapy by developing next generation image guidance for integration with radiotherapy planning. To achieve this, this project requires truly interdisciplinary expertise by combining a new Magnetic Resonance Imaging (MRI) method with deep learning based image analysis and segmentation tools and clinical translation. The ultimate aim is to enable a radiotherapy treatment plan precisely adapted to the individual patient’s tumor profile using “functional” MR imaging biomarkers. Integration of deep learning algorithms with the current clinical radiotherapy treatment software will allow the technique to be used for patients locally at our new Cancer Centre at Guy’s and St Thomas’ Hospitals.
Background
Radiotherapy is essential for cure or treatment of ~ 40% of all cancer patients in the UK. Image guidance plays a key role in delivering successful treatment. Currently a radiotherapy treatment plan is developed based on a CT scan. However, due to its superior soft tissue contrast MR is potentially far superior to guide successful treatment while minimising toxicity to adjacent healthy tissue. Further, the availability of “functional” MR contrasts enable to depict intra-tumoral characteristics such as diffusion, perfusion, oxygen status. Thus, the inclusion of MR in the radiotherapy workflow may prompt a paradigm shift towards next generation highly individualised radiotherapy by enabling focal dose escalation (“dose painting”), hypofractionation (fewer high precision, high dose treatments), voxelization (tissue sensitivity adapted dose distributions), adaptive radiotherapy (adaption of treatment after early response assessment), and motion control.
Key challenges
Integration of MR into RT goes beyond mere adaption of diagnostic MR protocols, but needs to address the following:
1) Achieve a fast and robust protocol to minimize the time patients are immobilized in the (uncomfortable) treatment position;
2) Acquire full FOV coverage to calculate a “pseudoCT” for accurate dosimetry;
3) Acquire 3D isotropic high-resolution and high contrast images to accurately delineate target and organs at risk (OAR).
Further, current radiotherapy treatment plans are manually created by segmenting target and organs at risk which is prone to error, time consuming and costly.
Specific Aims
In this proposal we aim to develop
1) RT-specific new prostate MRI protocol
a) Develop single-sequence multi-contrast MRI
b) Develop acceleration and motion correction
c) Develop multi-contrast functional biomarker MRI
2) Deep Learning (DL) integrated radiotherapy planning
a) Develop DL based “pseudo-CT” for accurate dosimetry
b) Develop DL based target and organ-at-risk delineation
PhD plan
Year 1) A RT-specific MRI sequence based on dual echo steady state (DESS) for multi-dimensional (3D spatial + “tunable” DESS contrast), full FOV, isotropic prostate imaging will be developed. To achieve robust and fast acquisition, the method will be extended to incorporate acceleration using variable density undersampling combined with iterative reconstruction and motion correction using integrated navigator data. Finally, to enable individualised adaptive radiotherapy guidance, the sequence will be further extended by integrating multi-contrast magnetisation preparation to encode RT-specific functional MRI biomarkers (DCE, diffusion, T2, T1, T2*).
Developments will be supported by developing an RT-specific full FOV geometry phantom, and healthy volunteer scanning.
Year 2) RT planning requires 1) accurate segmentation of the target and OARs and 2) photon attenuation information for accurate dosimetry. However, the current approach based on hand-drawn contours and involving image registration is time-consuming and may introduce unacceptable errors. We propose a novel dual-task neural network architecture to be combined with the single-sequence multi-contrast MR method developed in year 1. Dual-task learning is ideal, as it will jointly synthesize a pseudoCT scan (regression) and automated contours (segmentation). In a pilot patient study (n=15), both a clinical standard CT planning scan and the proposed MR-DL RT-plan will be acquired. We will evaluate the Mean Absolute Error and the fuzzy DICE score to evaluate segmentation between CT and MR-DL RT plans.
Year 3) The method will be evaluated in a pilot study in patients with prostate cancer. Contours and treatment plans derived from proposed method will be compared to current clinical approach. Dose to target and OAR toxicity will be compared. Towards clinical integration, in collaboration with medical physics and clinical fellow the developed algorithms will be implemented with the current clinical RT planning software (Eclipse).
Year 3.5) Final data analysis of clinical pilot study, publication and thesis write up.