Curvelet-Based Texture Classification in Computerized Critical Gleason Grading of Prostate Cancer Histological Images

摘要

Classical multi-resolution image processing using wavelets provides an efficient analysis of image characteristics represented in terms of pixel-based singularities such as connected edge pixels of objects and texture elements given by the pixel intensity statistics. Curvelet transform is a recently developed approach based on curved singularities that provides a more sparse representation for a variety of directional multi-resolution image processing tasks such as denoising and texture analysis. The objective of this research is to develop a multi-class classifier for the automated classification of Gleason patterns of prostate cancer histological images with the utilization of curvelet-based texture analysis. This problem of computer-aided recognition of four pattern classes between Gleason Score 6 (primary Gleason grade 3 plus secondary Gleason grade 3) and Gleason Score 8 (both primary and secondary grades 4) is of critical importance affecting treatment decision and patients’ quality of life. Multiple spatial sampling within each histological image is examined through the curvelet transform, the significant curvelet coefficient at each location of an image patch is obtained by maximization with respect to all curvelet orientations at a given location which represents the apparent curved-based singularity such as a short edge segment in the image structure. This sparser representation reduces greatly the redundancy in the original set of curvelet coefficients. The statistical textural features are extracted from these curvelet coefficients at multiple scales. We have designed a 2-level 4-class classification scheme, attempting to mimic the human expert’s decision process. It consists of two Gaussian kernel support vector machines, one support vector machine in each level and each is incorporated with a voting mechanism from classifications of multiple windowed patches in an image to reach the final decision for the image. At level 1, the support vector machine with voting is trained to ascertain the classification of Gleason grade 3 and grade 4, thus Gleason score 6 and score 8, by unanimous votes to one of the two classes, while the mixture voting inside the margin between decision boundaries will be assigned to the third class for consideration at level 2. The support vector machine in level 2 with supplemental features is trained to classify an image patch to Gleason grade 3+4 or 4+3 and the majority decision from multiple patches to consolidate the two-class discrimination of the image within Gleason score 7, or else, assign to an Indecision category. The developed tree classifier with voting from sampled image patches is distinct from the traditional voting by multiple machines. With a database of TMA prostate histological images from Urology/Pathology Laboratory of the Johns Hopkins Medical Center, the classifier using curvelet-based statistical texture features for recognition of 4-class critical Gleason scores was successfully trained and tested achieving a remarkable performance with 97.91% overall 4-class validation accuracy and 95.83% testing accuracy. This lends to an expectation of more testing and further improvement toward a plausible practical implementation.