Early esophagus cancer segmentation from gastrointestinal endoscopic images based on U-Net++ model

Abstract

Automatic segmentation of early esophagus cancer (EEC) in gastrointestinal endoscopy (GIE) images is a critical and challenging task in clinical settings, which relies primarily on labor-intensive and time-consuming routines. EEC has often been diagnosed at the late stage since early signs of cancer are not obvious, resulting in low survival rates. This work proposes a deep learning approach based on the U-Net++ method to segment EEC in GIE images. A total of 2690 GIE images collected from 617 patients at the Digestive Endoscopy Center, West China Hospital of Sichuan University, China, have been utilized. The experimental result shows that our proposed method achieved promising results. Furthermore, the comparison has been made between the proposed and other U-Net-related methods using the same dataset. The mean and standard deviation (SD) of the dice similarity coefficient (DSC), intersection over union (IoU), precision (Pre), and recall (Rec) achieved by the proposed framework were DSC (%) ​= ​94.62 ​± ​0.02, IoU (%) ​= ​90.99 ​± ​0.04, Pre ​(%) = ​94.61 ​± ​0.04, and Rec (%) ​= ​95.00 ​± ​0.02, respectively, outperforming the others. The proposed method has the potential to be applied in EEC automatic diagnoses.

Keywords

• Early esophagus cancer (EEC)
• Gastrointestinal endoscopic (GIE) images
• Semantic segmentation
• Supervised learning
• U-Net++