Radiotherapy: an AI outlines the organs to protect

Why does outlining take so long?

Radiotherapy treats cancer with beams aimed at the tumour. To plan it, the team first has to mark on a scan where the tumour is and where the neighbouring organs to protect are, called "organs at risk". This task is known as contouring.

It is done by hand, slice by slice, by specialist doctors. It takes time and varies from one expert to another and from one hospital to another, adding a heavy load to teams that are already busy.

How was the AI built and tested?

The researchers gathered 519 pelvic scans (prostate cancer) and 242 head and neck scans from 8 clinical sites. Fifteen organs at risk had been drawn by hand by experts, which gave the AI examples to learn from.

The models were trained on part of the data, then tested on internal scans and on scans from other hospitals that they had never seen during training. The team also compared the AI with the natural differences between several experts, and timed how long it took to correct or draw the outlines.

The scans were taken during treatment planning, from patients who received intensity-modulated radiation therapy between October 2013 and February 2020. The authors also stress that hand contouring introduces variability among both experts and institutions, which is one more reason to look for automatic help.

What were the results?

For 13 of the 15 structures, the AI reached an accuracy within the limits of the differences between experts. It worked as well on internal scans as on scans from other hospitals.

On time, correcting the AI's outlines took a mean of 4.98 minutes per scan for head and neck and 3.40 minutes for the prostate. Drawing the head and neck by hand took a mean of 86.75 minutes for an expert reader and 73.25 minutes for a radiation oncologist. The authors report a 93% reduction in time.

What could this mean for patients?

Less time spent drawing means more time for medical decisions and for preparing treatment. More uniform outlines from one hospital to another could also make planning more consistent.

The tool is designed to fit into existing work: the doctor stays in charge, checks every outline the AI draws and corrects it if needed before using it.

What are the limits of this study?

The scans come from 8 sites and cover only two types of cancer, prostate and head and neck. For two of the fifteen structures, accuracy was not within the range of expert differences. Correction time was measured on only 10 head and neck scans and 10 prostate scans.

Above all, the study measures outline quality and time saved, but does not show that patients get better health outcomes. That remains to be proven.