Automatic Computer-Aided Histopathologic Segmentation for Nasopharyngeal Carcinoma Using Transformer Framework

• Published in: Computational Mathematics Modeling in Cancer Analysis Vol. 13574; pp. 141 - 149
• Main Authors: Diao, Songhui, Tang, Luyu, He, Jiahui, Zhao, Hanqing, Luo, Weiren, Xie, Yaoqin, Qin, Wenjian
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Histopathological whole slide images; Nasopharyngeal carcinoma; Segmentation; Transformer
• ISBN: 9783031172656; 3031172655
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-17266-3_14

The segmentation of the histopathological whole slide images (WSIs) of nasopharyngeal carcinoma (NPC) plays an essential role in the diagnosis, grading and even prognosis analysis. Due to the huge size of pathological images and the fact that NPC often occurs in the middle and advanced stages, it is still challenging to generate accurate segmentation results automatically. Although many convolutional neural network (CNN) methods had achieved good segmentation performance in many types of images, however, the encoding of global context is insufficient, and it is prone to misjudge the adjacent regions. Meanwhile, the area of NPC pathological image is dense, which means that the image with a tiny size may fall into one category. To overcome this limitation, we apply a transformer-based framework on NPC pathological images that is designed for extracting and encoding global context information. To validate and compare the transformer framework with various CNN-based methods, experiments have been conducted on the clinical dataset collection of NPC. The transformer framework outperformed the state-of-the-art pure CNN-based methods in AUC and recall. Especially, our framework achieved 2.5%–3.5% higher DSC in 5X images and 2.1%–3.2% higher DSC in 10X images than other methods.