Cancer Cell Detection and Morphology Analysis Based on Local Interest Point Detectors

摘要

The automatic analysis of cancer cells has gained increasing relevance given the amount of data that biology researchers have to analyze. However, most biology researchers still analyze cells by visual inspection alone, which is time consuming and prone to induce subjective bias. This makes automatic cell image analysis essential for large scale, objective studies of cells. While the classic approach for automatic cell detection is to use image segmentation, in the case of in vivo brightfield images, such approach is not robust to image quality changes. To detect cells with robustness and increased performance we propose the use of local interest point detectors. We perform a comparison study between the use of the Laplacian of Gaussian filter, a Bank of Ring Filters and local convergence filters. Based on experimental results we found that the Laplacian of Gaussian filter outperformed all other in cell detection obtaining an accuracy of 78%. Additionally, through the analysis of shape fit, we found that the Laplacian of Gaussian filter obtained a better approximation to the shape of the cells having a Dice's coefficient of 81%.