Introduction

Head and neck cancer is the fifth most common cancer worldwide, with an annual incidence of approximately 500,000 cases globally. Radiotherapy is the primary non-surgical treatment of head and neck cancer and is commonly given in combination with chemotherapy and/or surgery. The aims of the treatment are to achieve loco-regional disease control whilst preserving organ function. Modern radiotherapy techniques allow the radiation dose delivered to the patient to be modulated in order to create highly conformal dose distributions, which minimise the doses delivered to normal tissues in close proximity to the tumour. However, there are still high rates of toxicity, which reduce patients’ quality of life and limit the amount of dose that can be delivered to the tumour and hence the probability of controlling the disease. A variation in radiotherapy-related toxicity between patients is observed. The most commonly reported severe radiation-induced toxicity during and following treatment is dysphagia (swallowing dysfunction). The radiation dose delivered to the mucosal lining of the throat (pharyngeal mucosa) is thought to be a major contributing factor to this side effect. Understanding the role of the radiotherapy dose distribution in the onset of dysphagia would allow new radiotherapy treatments aiming to reduce incidences of dysphagia to be designed.

Materials and methods

The 3D dose distributions delivered to the pharyngeal mucosa of 217 patients treated as part of several clinical trials in the UK were reconstructed. Pydicom was used to extract DICOM data (the standard file formats for medical imaging and radiotherapy data). NumPy and SciPy were used to manipulate the data to generate 3D maps of the dose distribution delivered to the pharyngeal mucosa. Several novel metrics to describe the dose distribution that encode spatial information were developed using NumPy and SciPy. Multivariate predictive modelling of severe dysphagia, including multiple descriptions of the dose distribution and relevant clinical factors such as comorbidities and concurrent treatments as covariates, was performed using Pandas, StatsModels and SciKit-Learn. MatPlotLib and Mayavi were used for plotting.

Results

An example of a map of the 3D dose distribution delivered to the pharyngeal mucosa is shown in figure 1 (see supporting document). The dose distribution was characterised by the volume, longitudinal and circumferential extents, and texture features at different dose levels as well as using 3D moment invariants. For predictive modelling a multivariate logistic regression model was chosen, as high model interpretability was deemed important for clinical use. Machine learning approaches for model feature selection were chosen due to high multicollinearity in the covariates. Model training, validation and testing are currently in progress.

Conclusions

A combination of Python modules has been used to generate 3D maps of the dose distribution delivered to the pharyngeal mucosa and characterise the dose distribution using spatial metrics. A multivariate logistic regression model using machine learning approaches for feature selection is being developed to predict severe dysphagia resulting from radiotherapy for head and neck cancer. It is hoped that this will lead to new radiotherapy treatments with lower incidences of severe dysphagia than the current technique.