Radiation therapy (RT) is an essential tool in loco-regional control of cancer:
of the roughly 1.7 million yearly cancer cases (> 15 million worldwide), an
estimated 50% can benefit for RT.  However there is still a great need to
improve the effectiveness of RT (overall survival, progression free survival)
while minimizing side effects.  The RT treatment planning process involves
delineation of the gross tumor volume (TV), which can vary among readers and
between trials of individual readers.  We are interested in studying and
exploiting the inter- and intra-observer variability when using CT, MR and PET
for delineation.  Two aspects are discussed.  We first demonstrate the
usefulness of MR and PET in delineating GTV.  We then describe the development
of automatic delineation methods that can predict radiologists' contours, using
deep learning, while taking into account contour variability.