To use radiation therapy to treat cancerous tumours, it is important to know exactly where the tumour is located. Oncologists, i.e. medical doctors with special training in diagnosing and treating cancer, also need to take organs at risk into account. An organ at risk is an important organ close to the tumour. By pinpointing the exact area, the radiation can eliminate the tumour tissue without damaging important adjacent tissue. To locate tumours and organs at risk, an MR scanner is used to collect different types of images. Before radiation therapy, a radiation oncologist must often manually draw the tumour and organs at risk in these images, which can take several hours for a single patient. By instead using artificial intelligence, AI, all areas can be drawn automatically and thus save many hours in healthcare.
Important with large training sets
AI models for radiotherapy are trained on a large number of medical images, where the radiation oncologist's drawings are used as the correct interpretation. It is important that the training set represents as large a percentage of the population as possible, so that the models can be used on all patients. However, there is a lack of medical images and medical information to train AI models with, as there are several legal aspects to consider. Among other things, ethical approval is required when such sensitive information is to be used for research.
Generating images to create large training sets
In an article published in the journal Scientific Data, LiU researchers describe how they evaluate the approach of instead generating images to train AI models with. Such images are called synthetic images.
“Synthetic images may be easier to share with other researchers. Since synthetic images do not belong to a specific person, GDPR is not applicable,” says senior associate professor Anders Eklund.
The researchers created the synthetic images using an open collection of data called BraTS. BraTS includes MR images and marked tumors for brain tumor patients from a variety of hospitals. Since the images are based on open data, the research team has been able to share the synthetic images on the AIDA (Analytic Imaging Diagnostics Arena) data hub, so that other researchers can use these images.
Berzelius helped them to save years of waiting
To evaluate the synthetic images, the research group compares the results from AI models trained with synthetic images with the results from AI models trained with real MR images of brain tumours.
“A challenge that remains is when we use too few real medical images to train the generative models. In these cases, we have seen that the synthetic images can be too similar to the real images, and we don’t want to share the real images”, says Anders Eklund.
Positive results
The results show that synthetic medical images can be an alternative to be able to share data more easily and thus develop AI models for medical applications. The exact legal conditions for sharing synthetic medical images remain to be investigated.
* The right answer to the image question can be found in Figure 5 in the article at nature.com
The dataset with the synthetic images shared by the research team
