Floris Reinders

UMC Utrecht

In short, Frank Hoebers' fellowship entails the following:

Transfer learning for auto segmentation of individual lymph nodes in patients with head and neck cancer

Most patients with head and neck cancer (HNC) treated with radiotherapy (RT) receive elective neck irradiation (ENI) to unilateral or bilateral lymph node levels for the treatment of possible, non-visible (occult) metastases. As literature shows that occult metastases are currently treated very well, it might be that conventional ENI protocols are currently too aggressive. With modern Magnetic Resonance Imaging (MRI) techniques it is possible to identify small individual lymph nodes inside these lymph node levels. Converting conventional ENI into specific dose delivery to individual lymph nodes will reduce radiation dose to healthy tissues and possibly reduce the long-term complication rate for patients with HNC without compromising regional control.
To enable irradiation of individual lymph nodes, all lymph nodes need to be contoured in the treatment planning CT/MRI. As patients with HNC contain approximately 30 lymph nodes on each side of the neck, manual contouring of these structures is very time consuming and will add approximately 2 additional hours of segmentation time. To reduce the extra workload, we have developed a convolutional neural network (CNN) in the University Medical Centre Utrecht (UMCU) for the automatic segmentation of individual lymph nodes on MRI images in HNC patients.

Although the performance our CNN is sufficient for clinical implementation, we expect worse outcomes if the CNN is applied on a new set of images in another medical center (due to different MRI scanners and scanning protocols). For the international implementation of individual lymph node irradiation we need fast and easy adaptation of the neural networks to the situation at hand at different medical centers.

In this study will examine if we can retrain our neural network on patient data acquired in Memorial Sloan Kettering Cancer Centre (MSKcc) in New York. The aim is to obtain a neural network that can run in MSKcc with a maximum of 5% reduction on all evaluation metrics compared to the UMCU CNN.