Skip to Main Content
Quantitative analysis of tumor cells is fundamental to pathological studies. Current practices are mostly manual, time-consuming, and tedious, yielding subjective and imprecise results. To understand the behavior of tumor cells, it is critical to have an objective way to count these cells. In addition, these counts must be reproducible and independent of the person performing the count. In this work, we propose a two-stage tumor cell identification strategy. In the first stage, potential tumor cells are segmented automatically using local adaptive thresholding and dynamic water immersion techniques. Unfortunately, due to histological noise in the images, a large number of false identifications are obtained. To improve the accuracy of the identified tumor cells, a second stage of feature rules mining is initiated. Experiment results show that image processing techniques alone are unable to give accurate results for tumor cell counting. However, with the use of features rules, we are able to achieve an identification accuracy of 94.3%.