A new approach to detect and segment overlapping cells in multi-layer cervical cell volume images
Hady Ahmady Phoulady, Dmitry B. Goldgof, Lawrence Hall, Peter R. Mouton · 2016
Detection and segmentation of overlapping cell nuclei and cytoplasm in multi-layer Pap smear volumes are critical prerequisites for automatic screening of cervical cancer. The novel approach proposed here is an improved version of our top ranked method for the Segmentation of Overlapping Cervical Cells from Multi-layer Cytology Preparation Volumes Challenge — ISBI 2015. From the images provided of multiple focal planes through liquid-based Pap smear samples, nuclei within Extended Depth of Field (EDF) images are first located with an iterative approach and then clumps are segmented using a learned Gaussian mixture model by Expectation-Maximization (EM) algorithm from pixel intensities. The overlapping boundaries of cytoplasm on the multifocal planes are approximated using a new distance metric and then refined in two post-processing steps. Promising results with Dice Coefficient (0.861), False Negative object rate (0.352) and True Positive pixel rate (0.874) indicate that nuclei and their corresponding cytoplasm in highly overlapping cytology multi-layer Pap smear volumes can be effectively detected and segmented using this approach.