An in silico imaging pipeline for the simulation and deep learning-based segmentation of photon-counting CT brain images

The combination of photon-counting CT and deep learning-based tumour segmentation can be of great benefit to radiotherapy treatment planning in neuro-oncology. We developed an in silico imaging pipeline for the simulation of photon-counting CT brain images. The reconstructed images are then used to train and test a 2D nnU-Net to automatically segment the tumour volumes and its performance is compared to conventional images acquired using an energy-integrating detector. Preliminary results demonstrate the capability of the pipeline to produce both photon-counting and conventional CT images, as well as images of different energy bins. We also demonstrate that the reconstructed photon-counting CT images are significantly more robust towards electronic noise as compared to the conventional CT images. Ongoing work focuses on the automatic tumour delineation and optimisation of image quality.