Recurrent Neural Networks for anatomy and dose nowcasting in radiotherapy

Objectives: Cancer accounts for >25% of deaths in the UK. A wide variety of therapies, such as chemotherapy and radiotherapy respectively, are available. Radiotherapy involves the delivery of radiation, targeted at the patient's disease, while aiming to minimise side effects caused by irradiating healthy tissue. Radiotherapy is delivered to >50% of cancer patients as part of their treatment. Currently, one radiotherapy plan designed on the patient's pre-treatment image is delivered throughout a patient's 4-6-week treatment. However, anatomy changes, and there is a need for full online adaptive radiotherapy to best fit the patient's anatomy each day. For some parts of the body where large day-to-day changes occur, e.g. cervix, bladder and abdominal tumours, this approach can reduce side effects.
To create a new plan each day, the standard radiotherapy workflow must be compressed from days to minutes. Adaptation must happen while the patient waits on the treatment couch: the longer this process takes the more likely their anatomy is to change, and any benefit of adaption is lost. Therefore, this project will devise methods to implement fully automatic, rapid safety checks using machine learning techniques to ensure the accuracy and safety of online adaptive radiotherapy.
Methods: Prior to treatment, patients will have a planning CT scan to define the target volume, neighbouring normal tissue and create a reference plan. This acts as a baseline for adapting the treatment to each day's anatomy. The longitudinal nature of a patient's dataset (i.e. images on subsequent days) naturally lends itself to the use of recurrent neural networks (RNNs) as a basis for ensuring the accuracy of the adaptation process.
RNNs have not been applied to medical image data before, therefore the optimal network architecture and training methodology needs to be explored. On the first day of treatment the network has no prior information, meaning its predictions will be population-based; later in the treatment process information about the individual patient is incorporated, creating a personalised prediction. Because of the transition from population-based to personalised prediction, quantifying the accuracy of the model is essential. Devising a robust methodology to handle longitudinal medical images within an RNN will form the basis of a methodological paper (paper 1).
The RNN will be applied to datasets where we have MR images of radiotherapy patients at multiple timepoints. (18x Cervical cancer patients at 4 time-points, 15x bladder cancer patients at 4 timepoints, 10 lung cancer patients at 3 timepoints, 10x pancreas cancer patients at 3 timepoints). The application of our RNN to a diverse set of treatment sites will show the generalisability of the approach and will form paper 2. Importantly, we will compare the speed of checking the classification of the anatomy with the RNN versus manual checks.
Next, an RNN will be applied to predict the radiation dose required to best treat the anatomy on a given treatment day. We will apply our RNN methodology to predict the optimal dose for each treatment day (using the initial reference plan + anatomy and any available subsequent days in the network). This prediction will form the basis for comparison against what the clinical adaption workflow creates. The datasets described above will be used for testing across multiple treatment sites (paper 3).
Finally, we will show proof-of-principle of a full clinical workflow (using generated data), illustrating how these models would work within the clinical environment and the potential for time-saving (paper 4).
Novelty: RNNs are commonly used in other fields, such as weather forecasting, but as yet have not been applied to the prediction of anatomy during radiotherapy treatment. This research will be the first application of RNN techniques in 3D volumetric data, and the first application of the techniques to the prediction of anatomy and dose.