AI, Image Analysis & Big Data
Truly personalised radiotherapy remains as challenging today as it did 20 years ago before the advent of: image guided radiotherapy; multi-leaf collimators; intensity-modulated radiotherapy (IMRT); and advanced (biological) imaging. Whilst each new generation of researchers has toiled with developing personalised radiotherapy strategies within their scope of practice, and available technology, the challenges faced in utilising the information/data available for true personalisation are unprecedented. What is more, tackling these challenges requires new approaches and solutions on a scale not seen before in the radiotherapy community.
The aim of our research is to address these challenges through integration of existing expertise and capabilities in a number of key radiotherapy research domains. These include: image analysis/artificial intelligence algorithm development; high-performance computing; clinical imaging; clinical trials; and importantly validation of new approaches. The backbone of our work, built-up over a number of years, is experience of developing artificial intelligence and image analysis algorithms for radiotherapy research and we now have a substantial software repository for this.
We are also part of a multidisciplinary research group combining expertise in Engineering, Chemistry, Physics, Medicine, Informatics and Computing to develop new adaptive radiotherapy strategies.
A new non-rigid image registration approach was developed to map cancer foci on MR images, where they are visible, to CT images where they are not visible. The method was used to investigate boosting the radiation dose to the tumour using a stereotactic ablative body radiotherapy (SABR) technique, which was found to be feasible and result in a reduction in rectal damage (Feng et al 2015).
Combining these artificial intelligence approaches to establish the location of the focal lesion perform a non-rigid mapping of the lesion between different imaging modalities, and time points, demonstrates the significant potential of this approach within an adaptive radiotherapy workflow.
In this pilot study on 57 non-small cell lung cancer (NSCLC) patients, pre-treatment radiotherapy planning CT images were analysed to identify patients at risk of pneumonitis after radiotherapy. Despite the fact that all patient CT images appear similar, and that no radiation had been given at the time of CT image acquisition, the accuracy of the approach for predicting pneumonitis was 87% when the imaging and clinical records data was combined (Montgomery et al 2017).
comparing radiomics and radiobiological approaches for predicting radiation-induced pneumonitis
The results below are preliminary results showing a comparison between radiomics and radiobiological approaches for predicting radiation-induced pneumonitis in non-small cell lung cancer patients.
Radiotherapy planning CT images and treatment plans from nine NSCLC patients made up of five patients that remained asymptomatic and four patients that showed signs of post-radiotherapy pneumonitis were analysed. A range of clinically available radiobiological models were used to find the normal tissue complication probability (NTCP) for each patient. A radiomics-based approach was used to calculate 99 texture analysis features on sub-regions of each image. These were classified using a Bayesian approach and a neural network (NN) to identify density patterns from asymptomatic and symptomatic patients.
Axial CT slices through the lung of a patient without radiation-induced pneumonitis (left) and a patient with radiation-induced pneumonitis (right). It is clear that both exhibit similar visual patterns, which are difficult to distinguish.
NTCP predictions for normal tissue in the left lung including the GTV.
NTCP predictions for normal tissue in the right lung including the GTV.
Texture analysis classification results obtained using a 32×32 subimage size.
Texture analysis classification results obtained using a 20×20 subimage size.