Saving time in the clinical workflow thanks to AI
Deep learning is currently revolutionising many research fields and can be used in health care to save a substantial amount of time in the clinical workflows. One example is radiotherapy treatment planning which requires 1) segmentation of the tumour to be killed by radiation, 2) segmentation of risk organs which should receive as little radiation as possible, and 3) generation of an optimal treatment plan from these segmentations. Deep learning can be used to save time, e.g. 10 – 60 minutes, in each of these steps. However, large, annotated datasets are required to train the complex AI models which usually contain millions of parameters.
In medical imaging, sharing sensitive data between hospitals is difficult due to ethics and regulations like GDPR. In addition, the number of patients with a rare disease in a specific city may be low. One way to enable large datasets in medical imaging is to train the AI models using federated learning, where the data stays at each hospital. In federated learning a powerful computer at each hospital uses the local image data to train the AI model (such as a segmentation network), and the computer at each hospital sends the updated local AI model to a combiner. The combiner aggregates the updates from each hospital and sends out a new global AI model to all hospitals from which the training continues.
In the ASSIST project, several partners in Sweden, the Netherlands and Belgium collaborated to train a brain tumour segmentation network using Swedish SME Scaleout’s FEDn framework for federated learning. The combiner was located in Uppsala, Sweden. Each partner used a unique part of an open dataset (BraTS) containing brain tumour images and annotations of each tumour. Future plans include improving the aggregation function, as well as using local radiotherapy treatment planning data from different cities.
Automation, Surgery Support and Intuitive 3D visualization to optimize workflow in IGT SysTems