PREDICTING THE LOCATION OF THE UTERINE CERVICAL OS FROM 2D IMAGES WITH CNN

Abstract

In an automated cervical cancer test, the prediction of the location of the cervical os from 2D images is required. Cervical os is the reference point to determine the lesion's location by either using cervical four-quadrant location or by 12 o’clock locations. Precise detection of the cervical os point ensures correct addressing of the lesions. This study used a 6-layer convolutional neural network to predict the center of the cervical os’ coordinates (x,y) on 2D grayscale images. We used a holistic approach without masking any visual element to predict the location of the cervical os. The 2D images were obtained using a telecentric lens and a CCD camera with light wavelengths of 500 550 nanometers. Because of the limited number of image samples (145 images), we used augmentation techniques to increase the training set size by rotating each original image in 1-degree increments from -30 degrees to +30 degrees relative to the center of the image. The 6-layer convolutional neural network was tested on 21 unseen cervix images using augmentation data. The outcomes showed that the image center-based augmentation technique improves the prediction performance. We obtained 2.4 RMSE in predicting the location of the cervical os.