Transductive prostate segmentation for CT image guided radiotherapy

Accurate 3-D prostate segmentation is a significant and challenging issue for CT image guided radiotherapy. In this paper, a novel transductive method for 3-D prostate segmentation is proposed, which incorporates the physician's interactive labeling information, to aid accurate segmentation, especially when large irregular prostate motion occurs. More specifically, for the current treatment image, the physician is first asked to manually assign the labels for a small subset of prostate and non-prostate (background) voxels, especially in the first and last slices of the prostate regions. Then, transductive Lasso (tLasso) is proposed to select the most discriminative features slice-by-slice. With the selected features, our proposed weighted Laplacian regularized least squares (wLapRLS) is adopted to predict the prostate-likelihood for each remaining unlabeled voxel in the current treatment image. The final segmentation result is obtained by aligning the manually segmented prostate regions of the planning and previous treatment images, onto the estimated prostate-likelihood map of the current treatment image for majority voting. The proposed method has been evaluated on a real prostate CT dataset including 11 patients with more than 160 images, and compared with several state-of-the-art methods. Experimental results indicate that the promising results can be achieved by our proposed method.