External beam radiation treatment for patients with cervical cancer is hindered by the relatively large motion of the target volume. A hybrid MRI-accelerator system makes it possible to acquire online MR images during treatment in order to correct for motion and deformation. To fully benefit from such a system, online delineation of the target volumes is necessary. The aim of this study is to investigate the accuracy of rigid, non-rigid and semi-automatic registrations of MR images for interfractional contour propagation in patients with cervical cancer. Registration using mutual information was performed on both bony anatomy and soft tissue. A B-spline transform was used for the non-rigid method. Semi-automatic registration was implemented with a point set registration algorithm on a small set of manual landmarks. Online registration was simulated by application of each method to four weekly MRI scans for each of 33 cervical cancer patients. Evaluation was performed by distance analysis with respect to manual delineations. The results show that soft-tissue registration significantly (P < 0.001) improves the accuracy of contour propagation compared to registration based on bony anatomy. A combination of user-assisted and non-rigid registration provides the best results with a median error of 3.2 mm (1.4-9.9 mm) compared to 5.9 mm (1.7-19.7 mm) with bone registration (P < 0.001) and 3.4 mm (1.3-19.1 mm) with non-rigid registration (P = 0.01). In a clinical setting, the benefit may be further increased when outliers can be removed by visual inspection of the online images. We conclude that for external beam radiation treatment of cervical cancer, online MRI imaging will allow target localization based on soft tissue visualization, which provides a significantly higher accuracy than localization based on bony anatomy. The use of limited user input to guide the registration increases overall accuracy. Additional non-rigid registration further reduces the propagation error and negates errors caused by small observer variations.