Examination Committee

Prof Volkan KURSUN, ECE/HKUST (Chairperson)
Prof Chin-Tau LEA, ECE/HKUST (Thesis Supervisor)
Prof Wai Ho MOW, ECE/HKUST
Prof Albert WONG, ECE/HKUST

Abstract

Cell segmentation is a critical task in automatic computer cytology diagnosis. However, overlapping cells pose a major challenge to cell segmentation because of blurred edges and inhomogeneous cytoplasm.

In this thesis, we first analyze the shape characteristic of cells by the Generalized Procrustes Analysis (GPA) and propose to exploit the obtained shape prior information by minimizing a nuclear norm in the active contour model.  Then we extend the classical active contour model to adaptively adjust the energy functional by introducing a dynamic parameter, which is derived using the Bayesian method.  Therefore, in the proposed method, different energies, which represent a variety of features of cell boundaries,   compete to eliminate the disturbances from nearby overlapping cells. Moreover, we develop an efficient algorithm to solve our model using the non-monotone Accelerated Proximal Gradient (nmAPG) method.

For our experiment, we design a complete framework for overlapping cell segmentation.  To evaluate the capability of our proposed method, we test it on the International Symposium on Biomedical Imaging (ISBI) 2014 and 2015 challenge datasets. Results demonstrate that our method significantly decreases the false negative rate (FNR) compared to the state-of-the-art methods.