Preliminary results from DeepMind’s collaboration with the Radiotherapy Department at University College London Hospital’s NHS Foundation Trust indicate that we are making considerable progress towards producing an artificial intelligence (AI) system that can undertake analysis and segmentation of medical scans regarding head and neck cancers to a similar standard as specialist clinicians. This procedure of segmenting is a critical but time-intensive phase when planning radiotherapy treatment. The discoveries also indicate that the framework can finish this procedure in a fraction of the time. 

Hastening the segmentation procedure 

Greater than half a million individuals receive diagnoses annually with cancers of the head and neck, globally. Radiotherapy is a critical part of treatment, but medical employees have to execute meticulous planning so that good tissue doesn’t get negatively impacted by radiation, a procedure which consists of radiographers, oncologists, and/or dosimetrists manually highlighting the regions of anatomy that require radiotherapy, and those regions which should be avoided. 

Even though the research is still at a nascent phase, the hope is that it could ultimately minimize the wait time between diagnosis and treatment, which could possibly enhance outcomes for cancer patients. The hope is that precise auto-segmentation could hasten the adaptive radiotherapy procedure, whereby radiotherapy treatments are adapted as the tumour gets smaller, even though more research is required to look into how this would function, practically speaking. 

In addition to having a positive impact on patient’s lives, health, and well-being, this work could also free up time resources for the practitioners who are treating them, implying they get to allocate additional time on patient care, education, and research. 

Application of the research to clinical environments 

Steps have been taken to make sure that the research is clinically viable and applicable. This consists of the development of a fresh performance metric and an evaluation set with new high-quality segmentations of scans chosen from sites prior unseen to the model which illustrates generalisability. These have both been open-sourced to the research community. But for the framework to have an influence on actual people who have received a cancer diagnosis, it requires expansion and demonstration that it functions well in clinical settings. 

This is why the researchers are looking ahead to move onto the next stage of work with UCLH, where a human evaluation of these artificial intelligence algorithms will be explored to evaluate how they could perform in a clinical setting. 

The spirit of collaboration is critical to success. Eventually, the belief is that sophisticated technologies will and ought to change lives, and it is thrilling to look forward to the next stages of this project.