Augmenting tumor sensitive matching flow to improve detection and segmentation of ovarian cancer metastases within a PDE framework

Abstract

The detection and segmentation of ovarian cancer metastases have potentially great clinical impact on women's healthcare. We recently developed a tumor sensitive matching flow (TSMF) technique to locate metastases by juxtaposing the roles of matching and classification within a PDE framework. This paper further augments the TSMF approach by integrating 1) shape index to measure metastasis-caused deformation, 2) Gaussian mixture model to describe metastasis intensity distribution, 3) total variation (TV) flow to enhance metastasis regions, and 4) TSMF vector displacements to control the amount of level-set propagation. The method was validated on contrast-enhanced CT data from 30 patients, of which 15 have 37 metastases in total. The true positive rate was 87% compared to 76% in our earlier work. Moreover, the false positive rate per patients was dropped to 1.1 from 4.2 in our earlier work. The metastasis segmentation achieved a Dice coefficient of 80.0±7.2%. All these experimental results demonstrated that shape index, Gaussian mixture model, TV flow, and TSMF-constrained level set propagation substantially improve the accuracy of metastasis detection and segmentation.