Contextual information-aided kidney segmentation in CT sequences

Abstract Based on the continuity of adjacent slices in a medical image sequence, a slice-based 3-D segmentation framework is constructed to extract the intact kidney by processing all slices automatically in the whole sequence. The framework includes four sections: initial segmentation, selection of the most reliable initial segmentation, location and modification of leakage. The crucial section of the proposed framework is selecting the most reliable initial segmentation image, which will be regarded as the reference image to evaluate the continuity of the following slice. Leakage location is carried out based on the contextual features, and the local iterative thresholding (LIT) is used to modify the leakage. As test examples of the framework, abdominal computed tomography (CT) images in enhanced phases are processed to segment kidney automatically. The total of 392 CT images in 7 sequences from 3 patients are selected as training images to determine the parameters in the database, and other 898 CT images in 21 sequences from 7 patients are used as test images to evaluate the effectiveness of the method. An average of three dimensional Dice similarity coefficient (3-D DSC ) of 94.7% and average symmetric surface distance ( ASSD ) of 0.91 mm are obtained, which indicate that the intact kidney can be perfectly extracted with hardly any leakage automatically.