Liquid-Based Pap Test Analysis Using Two-Stage CNNs

Abstract

According to the World Health Organization (WHO) cervix cancer is a real threat for women at earthly level. A practice to avoid those losses is an early diagnosis of the disease, generally done with the Papanicolaou or Pap test. This requires for a pathologist to check pap smear images in an arduous assignment, to determine the existence of suspicious or cancer cells. In third world countries doctors checks pap smear manually with microscopes, creating an enormous deficit of service. This paper proposes a TensorFlow ambient where the analysis of digital pap smears is carry out as a two-stage process. First, the sample is scanned using a ROI of $150\times 150$ pixels and two versions of the resulting image are stored in separated lists; one of low resolution ($20\times 20$ pixels) and one of high resolution ($250\times 250$ pixels). Then for the analysis, the first stage quickly evaluates the low-resolution images using a neural network that detects cells shapes saving their index (coordinates). In the second stage a specialized deep network uses this index to locate the high resolution images of the detected cells for zooming and recognition, being finally able to make high-resolution classifications. The software uses liquid-based pap smear equivalent to 460 patients with a 40x magnification. The trained system successfully classifies cells into normal and abnormal and could be big help to overloaded pathologists.