Segmentation of Cervical Images by Inter-subject Registration with a Statistical Organ Model

Abstract

For radiation therapy of cervical cancer, segmentation of the cervix and the surrounding organs are needed. The aim is to develop a fully automatic method for the segmentation of all relevant organs. Our approach is an atlas-based segmentation, with a registration scheme that is aided by statistical knowledge of the deformations that are to be expected. A statistical model that acts on the boundary of an organ is included as a soft constraint in a free-form registration framework. As a first evaluation of our approach, we apply it to the segmentation of the bladder. Statistical models for the bladder were trained on a set of manual delineations. Experiments on a leave-one-patient-out basis were performed, with the quality defined as the Dice similarity to the manual segmentations. Compared to a registration without the use of statistical knowledge, the segmentations are slightly, but significantly improved.