Clasificación de células en prueba de papanicolaou (pap test) en microscopía

Abstract

This short article addresses the classification of cervical cells in the Pap test, also known as the Papanicolaou test, using YOLOv7. The ultimate goal of the work is to achieve a tool that assists pathologists in the diagnosis of cervical cancer, improving quality and diagnostic times. Three experiments were performed to obtain preliminary results, using different class numbers from among the six existing classes according to the Bethesda system (Negative for intraepithelial lesions, ASC-US, ASC-H, LSIL, HSIL, SCC). The CRIC dataset was used to train the model, and a REST API in Flask and a web application were developed to use the trained models for inference. Preliminary results indicate that, with the six classes, the model fails to classify accurately enough, but is significantly improved by clustering the five positive classes. It is concluded that cervical cells are easily detectable with YOLOv7, and that better classification results could be obtained either by improving the object detection model, the dataset, or by introducing a convolutional classifier as a second step.