Deep neural network based assistive screening for cervical cancer

Harinarayanan K K, Nirmal Jith OU

Abstract:

Automating cervical cancer screening has the potential to reduce high mortality
rates due to cervical cancer, especially in developing countries. The most promising
of these techniques is assisted screening in which preliminary analysis is validated
by a pathologist. Assisted screening requires classification algorithms for initial
screening that is later validated by a pathologist. It also needs attention, detection
or segmentation algorithms for drawing attention of pathologists to important
regions. Existing algorithms for cervical cancer screening focus on classification
of individual cells . This focus leads to need for accurate segmentation of cells and
inability to use extracellular information. In this work we propose a segmentation
free deep learning algorithm for classification of PAP smear images. The proposed
algorithm uses the intrinsic information in the network to generate a map of
important regions for the pathologist to look into. This map is generated with sub
image resolution while the training data contains annotations at the image level
only. Our analysis on a dataset of around 14000 images validates the approach of
assisted screening in reducing the pathologist workload by a large factor.

Author Affiliation: Aindra Systems

Keywords: Cervical cancer, Papanicolaou test, Assistive screening