Automated Visual Examination (AVE) is a deep learning algorithm designed to enhance cervical precancer screening, particularly in low- and medium-resource regions. Trained on data from a comprehensive National Cancer Institute (NCI) study, AVE accurately identifies early-stage cervical neoplasia in uterine cervix images taken with a digital camera. It alerts healthcare providers when further evaluation is needed. However, for optimal results, AVE requires images of the cervix with specific quality criteria: sharp focus, good illumination, no shadows, and a complete view of the squamo-columnar transformation zone.
Previous efforts addressed some of these constraints. In this study, we introduce an innovative algorithm that determines the presence of the cervix in an image to an adequate degree. Inadequate or non-cervix images could yield incorrect results. Manually filtering such images is time-consuming, especially when dealing with large datasets lacking quality control.
Our solution leverages ensemble deep learning to differentiate cervix images from non-cervix images in smartphone-acquired cervical datasets. This ensemble method combines assessments from three deep learning architectures: RetinaNet, Deep SVDD, and a customized Convolutional Neural Network (CNN). Each architecture employs a distinct strategy—object detection, one-class classification, and binary classification—to make decisions.
We evaluate the performance of individual architectures and the ensemble on a separate test dataset comprising over 30,000 smartphone-captured images. The results are promising, with an average accuracy of 91.6% and an F-1 score of 0.890.
