Automated detection of anomalies in cervix cells using image analysis and machine learning

Abstract

Usage machine-based learning image cytometry to establish the diagnosis of cervix cancer using cellular morphology classification in comparison to the conventional cytological test. The study was divided into two phases consisting of 15 samples of cervix cells. In phase1, with previous diagnosis, the samples were divided into three groups of five samples each: normal (NC), low-grade squamous intraepithelial lesion (LGSIL or LSIL), and high-grade squamous intraepithelial lesion (HGSIL or HSIL). Images of cells were analyzed to create a training set of cells with known diagnosis for machine learning purposes. With the numerical data created, the software was trained to automatically classify the three types of cells. In phase 2, 885 cells were classified without previous diagnosis. In a last step, the classification of CPA was compared to cytopathology. NC and HSIL were identified with a high sensitivity and specificity (99%, 99%) and (98%, 97%) respectively. While the sensitivity and specificity of LSIL cells were lower (78%, 79%). It is possible to extract features of cervical cells by automatically generating numerical data that allowed the program to identify and classify different cell classes, using simple and low-cost reagents and free, reproducible softwires.