Retrospective Fat Suppression for Lung Radiotherapy Planning with Deep Learning Convolutional Neural Networks

We investigated three different Deep Learning techniques for performing retrospective fat suppression in T2 weighted imaging of lung cancer. The methods considered were two U-nets, using an L1 cost function or a conditional GAN, and a CycleGAN. The networks were trained on 900 images and then 16 test images were scored by 3 oncologists and a research radiographer. The L1 U-net and CycleGAN were scored at 73% and 72% respectively, relative to a gold standard of 80% for prospectively fat saturated images, and the scorers indicated they would be happy to use the generated images for radiotherapy target delineation.