Nuclei Segmentation and Identification in Practical Pap-Smear Images with Multiple Overlapping Cells

摘要

Uterine cervical cancer is the second most common cancer among women, with around half a million cases reported every year. The Papanicolaou technique provides an inexpensive early diagnosis if adequate Pap-smear images are available. A system was developed to automatically capture hundreds of Pap-smear images from slide samples provided from a rural location. The system selects images that present a sufficient number of cell nuclei and signs of abnormal cells and sends them to a remote hospital in an urban area for specialized diagnosis. This paper focuses on a technique to identify and segment cell nuclei from multiple overlapping cervical cells. The proposed method first segments and selects nucleus candidates based on a novel split and merge-search process. Splitting by mean-shift over-segmentation is followed by selection of candidate regions based on the application of basic a priori restrictions to an incremental region-merging process inspired by level sets. Each candidate is then described by shape, color, and spatial-context features to train and test several classification schemes. The identification results include over 99.3% recall and over 99.2% precision in 2-fold cross validation, and the segmentation results are comparable to those obtained by a specialist.