Automatic localization of solid organs on 3D CT images by a collaborative majority voting decision based on ensemble learning

Abstract

Purpose: Organ segmentation is an essential step in the development of computer-aided diagnosis/surgery systems based on computed tomography (CT) images. A universal segmentation approach/scheme that can adapt to different organ segmentations can substantially increase the efficiency and robustness of such computer-aided systems. However, this is a very challenging problem. An initial determination of the approximate position and range of a target organ in CT images is prerequisite for precise organ segmentations. In this study, we have proposed a universal approach that enables automatic localization of the approximate position and range of different solid organs in the torso region on three-dimensional (3D) CT scans.

Methods: The location of a target organ in a 3D CT scan is presented as a 3D rectangle that bounds the organ region tightly and accurately. Our goal was to automatically and effectively detect such a target organ-specific 3D rectangle. In our proposed approach, multiple 2D detectors are trained using ensemble learning and their outputs are combined using a collaborative majority voting in 3D to accomplish the robust organ localizations.

Results: We applied this approach to localize the heart, liver, spleen, left-kidney, and right-kidney regions independently using a CT image database that includes 660 torso CT scans. In the experiment, we manually labeled the abovementioned target organs from 101 3D CT scans as training samples and used our proposed approach to localize the 5 kinds of target organs separately on the remaining 559 torso CT scans. The localization results of each organ were evaluated quantitatively by comparing with the corresponding ground truths obtained from the target organs that were manually labeled by human operators. Experimental results showed that success rates of such organ localizations were distributed from 99% to 75% of the 559 test CT scans. We compared the performance of our approach with an atlas-based approach. The errors of the detected organ-center-positions in the successful CT scans by our approach had a mean value of 5.14 voxels, and those errors were much smaller than the results (mean value about 25 voxels) from the atlas-based approach. The potential usefulness of the proposed organ localization was also shown in a preliminary investigation of left kidney segmentation in non-contrast CT images.

Conclusions: We proposed an approach to accomplish automatic localizations of major solid organs on torso CT scans. The accuracy of localizations, flexibility of localizations of different organs, robustness to contrast and non-contrast CT images, and normal and abnormal patient cases, and computing efficiency were validated on the basis of a large number of torso CT scans.