General Description of the Problem

The automated detection and segmentation of overlapping cells using microscopic images obtained from Pap smear [1] can be considered to be one of the major hurdles for a robust automatic analysis of cervical cells. The Pap smear is a screening test used to detect pre-cancerous and cancerous processes, which consists of a sample of cells collected from the cervix that are smeared onto a glass slide and further examined under a microscope. The main factors affecting the sensitivity of the Pap smear test are the number of cells sampled, the overlap among these cells, the poor contrast of the cell cytoplasm, and the presence of mucus, blood and inflammatory cells [2]. These factors all exacerbate intra- and inter-observer variability and lead to a large variation in false negative results [3]. These issues have motivated the development of automated cell deposition and automated slide analysis techniques. Automated cell deposition techniques, such as mono-layer preparations, remove a large portion of blood, mucus and other debris, reduce cell overlap and produce cells that are more likely to occur in a single focal plane. This makes both manual and automated slide analysis faster and easier [4]. Automated slide analysis techniques attempt to improve both sensitivity and specificity by automatically detecting, segmenting and classifying the cells present on a slide [5, 6, 7, 8, 9].

The main focus of this challenge is on cell detection and cell segmentation for the automated analysis of cervical cytology specimens. Current systems can segment the nucleus and cytoplasm of cervical cells in isolation [7] (i.e., cells without any overlap with other cells), segment overlapping nuclei [10,11,12] and segment overlapping nuclei plus the whole region representing the cellular clumps [6,2]. Only recently has the complete segmentation of overlapping cells been addressed [13,14], but these methods show results that are still not robust enough for clinical practice. Indeed, the effectiveness of these methods is severely complicated by the fact that several layers of cervical cells are present in a glass slide (see Fig.1), which means that cells in an upper layer can partially occlude cells lying underneath [12]. In a manual examination, the user can adjust focus in order to assist interpretation of the overlapping cells, using the depth cue focus provides. Given the difficulty in automating this focus adjustment, we propose the use of an extended depth of field (EDF) cytology image that puts all objects in focus in a single image [15]. This is much simpler and more efficient than trying to separate the transparent layers from different focal depths, e.g., 16, but the use of these EDF images means that the methodology must analyse overlapping cells that are all in focus in a single image. Therefore, we propose this challenge on the detection and segmentation of overlapping cells based on the analysis of individual fields of view obtained from a mono-layer cytology preparation.