AGMC-TU Pap-Smear Cytological Image Dataset: Creation, Annotation, and Analysis towards Early Detection of Cervical Cancer

Automatic cervical cancer screening based on pap-smear images is a highly effective tool where the cells are categorized into normal and abnormal. However, success of most automation tool depends on the accurate extraction of features from the pap-smear images that represent some discriminative characteristics between these two categories of cells. In this paper, we described the designing protocols for creation of a new pap-smear image dataset entitled as AGMC-TU Pap-Smear Cytological Image Dataset. The dataset comprises of 50 normal and 50 abnormal pap-smear images belonging to ethnic and non-ethnic populations of low resource cervical cancer prone regions. Moreover, ground truths of suspicious nucleus regions are annotated in terms of pixel oriented binary masks are also provided with the dataset. Analysis of our dataset includes a conventional (i.e., shape features) and deep feature based study of pap-smear images by dividing them into two major groups: normal and abnormal. Outcome of the analysis clearly differentiates normal and abnormal pap-smear images.