A Learning based Hierarchical Framework for Automatic Prostate Localization in CT Images

Abstract

Accurate localization of prostate in CT images plays an important role in image guided radiation therapy. However, it has two main challenges. The first challenge is due to the low contrast in CT images. The other challenge is due to the uncertainty of the existence of bowel gas. In this paper, a learning based hierarchical framework is proposed to address these two challenges. The main contributions of the proposed framework lie in the following aspects: (1) Anatomical features are extracted from input images, and the most salient features at distinctive image regions are selected to localize the prostate. Regions with salient features but irrelevant to prostate localization are also filtered out. (2) An image similarity measure function is explicitly defined and learnt to enforce the consistency between the distance of the learnt features and the underlying prostate alignment. (3) An online learning mechanism is used to adaptively integrate both the inter-patient and patient-specific information to localize the prostate. Based on the learnt image similarity measure function, the planning image of the underlying patient is aligned to the new treatment image for segmentation. The proposed method is evaluated on 163 3D prostate CT images of 10 patients, and promising experimental results are obtained.

Fig. 1

(a) An image slice obtained from a 3D prostate CT image. (b) The prostate boundary highlighted by the red manually-delineated contour superimposed on the image in (a). (c) An image slice with bowel gas obtained from a different 3D prostate CT image of the same patient as in (a).

Fig. 2

Examples showing some selected distinctive regions. The selected regions are highlighted in green. The prostate boundary determined by the clinical expert is delineated in red.

Fig. 3

Centroid distance between the estimated prostate volume and the manually segmented prostate volume (a) without using the online update mechanism, and (b) with the online update mechanism. The horizontal lines in each box represent the 25th percentile, median, and 75th percentile respectively. The whiskers extend to the most extreme data points.

Fig. 4

typical performance of the proposed method on six slices of the 14th treatment image of the 8th patient. The Dice ratio of this treatment image is 90.42%. The black contours show the results of manual segmentation, and the white contours show the results obtained by the proposed method.