Sensors & Transducers (Feb 2021)
The Automated Diagnosis Architecture and Deep Learning Algorithm for Cervical Cancer Cell Images
Abstract
Cervical cancer is the second most common cancer in women worldwide, with approximately 500,000 cases each year. Approximately 80 % of the reported deaths are observed in countries with poor medical conditions, such as developing countries; 230,000 people die each year from cervical cancer. When detected early, it is possible to prevent progression to invasive cancer by adequate treatment. Therefore, it is very important to detect human papillomavirus (HPV), which is known as a major cause of cervical cancer, and cervical cancer that has already progressed. The most well-known test to diagnose cervical cancer is the Pap test. Although the Pap test consumes more time to diagnose cervical cancer, this test has saved the lives of many patients through early detection of cells progressing into cervical cancer for treatment and has contributed significantly to the prevention of cervical cancer and the reduction of mortality due to cervical cancer. However, since the Pap test requires emanating the cell slide of each patient, which takes a lot of time, the pathologists performing Pap tests tend to become very tired, and it is very difficult to examine many patients. Therefore, the need for diagnostic aids to help pathologists make quick decisions is on the rise. This study aimed to address this problem by developing an automatic diagnostic aid tool for cervical cancer using Yolo V3, a deep learning algorithm. First, the RGB cell image is converted into a gray-scale image by pre-processing, and the noise in the image is removed using a 2-dimensional Gaussian smoothing filter. Next, each cell image for training was labeled by a pathologist. Finally, using the trained algorithm, the cells in the image from the Pap test are identified and marked by bounding boxes to aid the pathologist with rapid diagnosis. For training of the model, 5,631 pre-processed Pap test images were used, and the model was then tested using 563 images. In this process, the performance indicator of PASCAL Visual Object Classes was used, which was set by raising the threshold from 0 to 1. The precision and recall values for all test images were obtained to calculate the average precision value corresponding to the recall value and to use it as an indicator to determine the performance of the algorithm. The average precision in this study was 73.34 %, and the model may be used as an auxiliary tool for pathologists performing Pap tests by improving accuracy using additional data in the future.
