Cellular quantitative analysis of neuroblastoma tumor and splitting overlapping cells
1 Centre for Quantum Computation and Intelligent Systems (QCIS), Faculty of Engineering and IT, University of Technology, Sydney, PO Box 123, Broadway, NSW 2007 Sydney, Australia
2 Head of Biospecimens and Tumor Bank, Children’s Cancer Research Unit, The Kids Research Institute, The Children’s Hospital at Westmead, Locked Bag 400, Westmead, NSW 2145 Sydney, Australia
3 QCIS and the Head of School of Software in the Faculty of Engineering and IT, University of Technology, Sydney, PO Box 123, Broadway, NSW 2007 Sydney, Australia
BMC Bioinformatics 2014, 15:272 doi:10.1186/1471-2105-15-272Published: 11 August 2014
Neuroblastoma Tumor (NT) is one of the most aggressive types of infant cancer. Essential to accurate diagnosis and prognosis is cellular quantitative analysis of the tumor. Counting enormous numbers of cells under an optical microscope is error-prone. There is therefore an urgent demand from pathologists for robust and automated cell counting systems. However, the main challenge in developing these systems is the inability of them to distinguish between overlapping cells and single cells, and to split the overlapping cells. We address this challenge in two stages by: 1) distinguishing overlapping cells from single cells using the morphological differences between them such as area, uniformity of diameters and cell concavity; and 2) splitting overlapping cells into single cells. We propose a novel approach by using the dominant concave regions of cells as markers to identify the overlap region. We then find the initial splitting points at the critical points of the concave regions by decomposing the concave regions into their components such as arcs, chords and edges, and the distance between the components is analyzed using the developed seed growing technique. Lastly, a shortest path determination approach is developed to determine the optimum splitting route between two candidate initial splitting points.
We compare the cell counting results of our system with those of a pathologist as the ground-truth. We also compare the system with three state-of-the-art methods, and the results of statistical tests show a significant improvement in the performance of our system compared to state-of-the-art methods. The F-measure obtained by our system is 88.70%. To evaluate the generalizability of our algorithm, we apply it to images of follicular lymphoma, which has similar histological regions to NT. Of the algorithms tested, our algorithm obtains the highest F-measure of 92.79%.
We develop a novel overlapping cell splitting algorithm to enhance the cellular quantitative analysis of infant neuroblastoma. The performance of the proposed algorithm promises a reliable automated cell counting system for pathology laboratories. Moreover, the high performance obtained by our algorithm for images of follicular lymphoma demonstrates the generalization of the proposed algorithm for cancers with similar histological regions and histological structures.