An Empirical Study of Vision Transformers for Cervical Precancer Detection

• Published in: Recent Trends in Image Processing and Pattern Recognition Vol. 1576; pp. 26 - 32
• Main Authors: Angara, Sandeep, Guo, Peng, Xue, Zhiyun, 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: Cervical cancer; Transfer learning; Vision Transformers (ViT)
• ISBN: 3031070046; 9783031070044
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-3-031-07005-1_3

Cervical precancer is a direct precursor to invasive cervical cancer and a prime target for ablative therapy. This paper presents an empirical study of Vision Transformers (ViT) for cervical precancer classification, an extended study of our previous work using data derived from two studies conducted by the U.S. National Cancer Institute. In this study, we show that ViT can significantly outperform the current state-of-art methods. We also examine data augmentation techniques that help reduce noise that can interfere in precancer detection, such as specular reflection. We achieve 84% accuracy on the test set outperforming the existing works based on the same dataset. Apart from the performance gains, we observe the learned features focus on cervical regions of anatomical significance. Through these experiments, we demonstrate that ViT attains excellent results compared to the current state-of-the-art methods in classifying cervical images for cervical precancer screening.