Requirement of an Automated system for primary screening cervical smears

In 1984 the International Academy of Cytology defined the requirements

the system must not pass as negative any sample that contains malignant cell
the system should be able to detect artifacts and identify infectious organisms such as trichomonas vaginalis and candida species

Current requirements are more modest. They are :

The system should to select suspicious cells (or slides) and present them to the cytologist for final classification
Smears excluded from human evaluation should not contain abnormalities at a rate exceeding that acheivable with primary manual screening.
that the system should be able to comment on smear adequacy, scanning should be rapid and reliable and reproducible.
the system should not exclude Pap stained smears suitable for classification
the sensitivity of the automated device (plus the cytologist) should equal or excede the sensitivity of the =primary screener (ie 95% sensitivity for HSIL).

Tasks required of the computer to detect recognise and classify relevant cells in cervical smears

In order to ensure the detection , recognition and classification of relevant cells in the smear the computer is required to perform a number critical tasks

Load a slide onto the motorised microscope stage
Find a starting point for scanning and proceed to scan the slide in a predetermined way
Find objects of interest and determine their focal plane
Detect cervical epithelial cells from among other objects on the slide
Label each epithelial cell and define its boundaries and recognise nucleus and cytoplasm (a process known as segmentation)
-Extract diagnostic features of the nucleus eg chromatin content and structure and cytoplasm (over 100 geometric or texture features may be extracted)
Eliminate irrelevant cells, debris, artifacts, overlapping and touching cells
Finally classify the cells as normal or abnormal

Requirements of image analysis systems

Many of the early automated devices relied entirely on image analysis of the smear on a cell by cell basis using an algorithmic (rule based) computer software programme. This approach failed initially largely because insufficient computer power to process the huge computational load required for the analysis of a cervical smear in this way.

A recent development in computer technology for the automation of cervical screening has been the introduction of artificial intelligence systems of image processing into the automated analysis of cervical smears. These systems include neural network systems which represent a non algorithmic branch of artificial intelligence which is particularly suited to recognising patterns in natural scenes and as such is particularly appropriate for the analysis of cervical smears where much of what the screener recognises as abnormality cannot be precisely defined in purely objective or morphometric terms. However, in contrast to the algorithmic programmes, neural networks are rather slow.

The current commercial automated systems utilise both algorithm based and artifical intelligence programs for the analysis of cervical smears