Cleaning Highly Unbalanced Multisource Image Dataset for Quality Control in Cervical Precancer Screening

• Published in: Recent Trends in Image Processing and Pattern Recognition Vol. 1576; pp. 3 - 13
• Main Authors: Xue, Zhiyun, Guo, Peng, Angara, Sandeep, Pal, Anabik, Jeronimo, Jose, Desai, Kanan T., Ajenifuja, Olusegun K., Adepiti, Clement A., Sanjose, Silvia D., Schiffman, Mark, Antani, Sameer
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Communications in Computer and Information Science
• Subjects: Acetowhitening; Anomaly detection; Cervical cancer; Cross-dataset evaluation; Deep learning; Green-filter; Highly unbalanced dataset; Lugol’s iodine
• ISBN: 3031070046; 9783031070044
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-3-031-07005-1_1

Automated visual evaluation (AVE) of uterine cervix images is a deep learning algorithm that aims to improve cervical pre-cancer screening in low or medium resource regions (LMRR). Image quality control is an important pre-step in the development and use of AVE. In our work, we use data retrospectively collected from different sources/providers for analysis. In addition to good images, the datasets include low-quality images, green-filter images, and post Lugol’s iodine images. The latter two are uncommon in VIA (visual inspection with acetic acid) and should be removed along with low-quality images. In this paper, we apply and compare two state-of-the-art deep learning networks to filter out those two types of cervix images after cervix detection. One of the deep learning networks is DeepSAD, a semi-supervised anomaly detection network, while the other is ResNeSt, an improved variant of the ResNet classification network. Specifically, we study and evaluate the algorithms on a highly unbalanced large dataset consisting of four subsets from different geographic regions acquired with different imaging device types. We also examine the cross-dataset performance of the algorithms. Both networks can achieve high performance (accuracy above 97% and F1 score above 94%) on the test set.