Clinical Applicability of Deep Learning for Organ-At-Risk Segmentation in Radiotherapy Planning

Cancer is one of the most common ailments of humanity and the second most prominent death cause in the developed world. With the predicted increase in the cancer burden, there is a rising need for automation in cancer therapy. Radiotherapy is one common treatment modality and requires the delineation of anatomical structures for planning purposes. With the rise of deep learning, new automatic contouring methods have been developed. This thesis aims at providing evidence for the validity of these methods in clinical practice and investigates the usability of third-party provided tools. The research was conducted by using large multi-institutional data sets for the female breast and pelvis as well as the male pelvis. Further, two multi-institutional studies were performed to evaluate the quality of a deep neural network for the segmentation of organs-at-risk of the female breast and the male pelvis. Further, the generalizability of a deep neural network towards patients from other hospitals was evaluated. We show that the clinical acceptability of segmentation by the deep neural network for the male pelvis is equivalent or better than for segmentations stemming from clinical practice. Further, we show that the average contouring time for the delineation of both breasts and the heart can be reduced from approximately 20 min to 3 min by using the developed deep neural network. Additionally, we show that a third-party provided model for female breast, female pelvis and male pelvis can readily be used for structures with well defined anatomical borders, while a hospital-specific model performs better for the breasts. This thesis gives evidence that deep neural networks can be used in a clinical setting and, with a few exceptions, deep neural networks from a third-party provider can readily be adopted by a hospital.