#3 Research Meet-up: Analytics in the fight against cancer: 30 million patients treated with IMRT

Around one in three women and one in two men will develop cancer during their lifetime. The project from this meetup explores how analytics is contributing to radiation therapy, from identifying where treatment should be directed to optimising radiation doses and adapting treatment to an individual patient’s response.

Why does analytics matter in cancer treatment?

In several countries, cancer has surpassed cardiovascular disease as the most likely cause of death. At the same time, treatment outcomes have improved substantially: the five-year survival rate after a cancer diagnosis has increased from around 50% in the 1970s to 70% today, while the age-standardised cancer death rate has fallen by 15% globally since 1990.

This meet-up focused on three analytical challenges in cancer diagnosis and treatment:

1. locating the tumour;

2. delivering treatment to the tumour while limiting radiation to surrounding healthy organs;

3. individualising treatment, including deciding when treatment should stop.

The discussion centred on radiation therapy, one of the main forms of cancer treatment and a field in which treatment planning depends heavily on quantitative methods.

How can analytics define where radiation should be delivered?

Automated identification, or segmentation, of tumours and healthy organs is increasingly becoming part of clinical practice through advances in convolutional neural networks, a type of machine-learning model used for analysing images.

One problem remains difficult: automatically defining the clinical target volume (CTV). The CTV includes both the tumour visible on a scan and microscopic cancer extensions that cannot be seen directly using diagnostic imaging.

Physicians currently draw these areas manually on CT or MR images. This introduces uncertainty and can lead to differences between clinicians, as well as differences when the same clinician repeats the task.

One direction is therefore to model how microscopic disease may extend beyond the visible tumour and quantify the uncertainty around it. The lecture will discuss the use of shortest-path algorithms, which estimate possible routes of disease extension from the visible tumour while respecting anatomical barriers, as an initial step towards this goal.

How can optimisation target the tumour while limiting radiation to healthy organs?

Analytics has had its largest impact in radiation therapy through the development of optimisation methods for intensity-modulated radiation therapy (IMRT).

IMRT allows the intensity of incoming radiation beams to vary. The planning problem is to determine the beam intensities that provide the required radiation dose to the CTV while reducing exposure to surrounding organs at risk (OARs) — healthy organs that could be damaged by radiation.

This is an inverse planning problem: clinicians specify the treatment goals, and optimisation methods calculate radiation patterns that can achieve them. The mathematical challenge has similarities to reconstructing an image in computed tomography.

The meet-up covered several ways of formulating this problem, including multi-objective optimisation, where competing goals such as tumour coverage and protection of several organs need to be balanced.

The models can include straightforward minimum and maximum dose limits as well as more complex dose-volume constraints, which restrict how much of an organ may receive a specified radiation dose. Mathematically, these resemble value-at-risk constraints used in financial planning.

Some treatment requirements are expressed in terms of radiation dose, while others concern the intensity of the radiation beams. Split feasibility algorithms can move between these different mathematical spaces to identify plans that satisfy requirements in both.

How does an optimised plan become a treatment that can actually be delivered?

An optimised radiation profile also needs to be translated into physical movements of the treatment machine.

The lecture will discuss how IMRT intensity profiles are converted into beam shapes that can be delivered using multileaf collimators — devices containing many individually controlled metal leaves that shape a radiation beam.

It will also consider volumetric modulated arc therapy, where the treatment machine's gantry rotates continuously around the patient while the radiation beam changes shape and intensity.

Optimisation research has been central to making these approaches practical. Since IMRT was first introduced clinically in the 1990s, approximately 30 million patients have been treated using the technique. Clinical examples will illustrate how these methods are used in practice.

What can analytics mean for personalised cancer treatment?

The final part of the meet-up looked towards personalised medicine.

Rather than determining an entire treatment course in advance and applying the same plan regardless of how an individual responds, future analytical approaches could use information collected during treatment to adapt subsequent decisions.

One question is optimal stopping: determining when continuing treatment is likely to remain beneficial for an individual patient and when treatment should be changed or stopped.

This extends the role of analytics in radiation therapy from designing the treatment plan itself towards supporting treatment decisions as an individual patient’s response becomes known.