Detecção automática de células cervicais com base em atributos radiais

Abstract

Microscopic quantification of cervical cell properties has been used for the early detection of precancerous lesions from Pap smears for decades. The traditional approach relies on the visual screening of a Pap smear glass slide to search for patterns correlated to abnormal cells. The major challenge in cervical cell screening for public health programs is the reliance upon manual inspection by different pathologists, a task that does not scale to the population growth. This work introduces cell classification and image retrieval computational tools, including a new radial feature description (RFD) to distinguish normal and abnormal patterns. The key idea lies in defining evenly interspaced segments around the cell nucleus, and proportional to the convexity of the nuclear boundary. The main advantage of the proposed RFD is the sensitivity to the intensity variation around the nuclear membrane, without cytoplasm outlining, and combining chromatin distribution through texture features. For performance evaluation, we applied two databases: the Herlev and a higher resolution one called BHS, both with thousands of samples. We create the BHS database by applying a new nucleus segmentation method that we proposed here. Then, we classify cells with Random Forest and bootstrap, and perform content-based image retrieval with the cosine similarity, comparing our methodology to other cell recognition techniques. The main contributions are: a) a new method for cervical cell nuclei segmentation; b) RFD as a fast and an accurate cervical cell descriptor, without prior cytoplasm segmentation; b) the BHS database with high-resolution images and ground-truth; c) pyCBIR, a new tool for image retrieval; d) classification and CBIR results using 14 different algorithms including the proposed RFD and two convolutional neural networks. Our results show that the proposed RFD allows accurate discrimination between normal and abnormal cervical cells, achieving the highest accuracy measures in terms of Kappa for both Herlev and BHS cell image sets.