Integrating an adaptive region-based appearance model with a landmark-free statistical shape model: application to prostate MRI segmentation