Precise contouring of organs is an important but time-consuming part of radiotherapy treatment planning. Every patient requires an individual approach. The Cloud Atlas project developed an auto-contouring solution taking advantage of the latest innovations in machine learning.
Research shows that some 80% of clinical institutions already have some sort of automatic contouring capability, but only 30% use it routinely as the contour editing process does not gain any time for the clinical teams. Cloud Atlas initially looked to develop a bigger database, hosted in the Cloud, to yield better results.
The original idea of the project was to collate existing contour data from a range of hospitals and existing databases (or atlases) to provide a potential improved starting point for contouring in new patients. Precise contouring of cancer tumours and surrounding organs is essential to ensure that radiotherapy treatment is accurately targeted and damage to adjacent tissue is minimised.
“However, half way through it became clear that there were issues with our original assumptions for the project,” says Dr Mark Gooding, Chief Scientist at Mirada Medical. “Accessing useful patient data was proving challenging and the process was not improving significantly in terms of efficiency, so it was decided to pivot the project and see if deep learning techniques and artificial intelligence (AI) could make a difference.”
A deep learning approach used data from some 500 patients to train on. The project developed an auto-contouring solution based on artificial neural networks, together with a web-based viewing application for radiotherapy target objects. These results were then subjected to clinical validation via project partner the Maastro Clinic in the Netherlands and compared to manual contouring results, as well as those provided by state of the art automatic atlas-based contouring.
The deep learning approach was found to provide high quality contouring and reduced the average time required to refine and complete contouring for patients.
“This is very exciting for us,” says Gooding. “There is a lot of hype around AI, but the deep learning approach has delivered a useful clinical product – one that answers a real clinical need.”
By providing a better basis for individual contouring the Cloud Atlas system frees up time for clinicians to plan the actual dosimetry programme and maximise the effectiveness of the treatment. This results in improved healthcare through reduced time required to plan a treatment overall, reduced costs associated with that planning, and improved overall consistency of the contouring process.
The new solution has also effectively established a new ‘clinical’ Turing test. When clinicians were asked to determine if a particular contour was drawn by a human or by the AI-based solution, they found this difficult. “Doctors still felt the need to edit contours produced by the software, but they found that any errors made were often the sort made by human colleagues,” observes Gooding.
“The Cloud Atlas project shows that AI is starting to deliver what professionals need,” he concludes.
The Cloud Atlas solution was introduced to a number of potential early adopters at the beginning of 2018 and was given a full commercial launch at the 37th Annual Congress of the European Society for Radiotherapy and Oncology (ESTRO) on 20 to 24 April 2018 in Barcelona, Spain.