Computer Aided Annotation of Early Esophageal Cancer in Gastroscopic Images based on Deeplabv3+ Network

Abstract

The diagnoses of Early Esophageal Cancer (EEC) based on gastroscopic images is a challenging task in clinic, which relies heavily on subjective artificial detection and annotation. As a result, computer aided diagnosis (CAD) methods that support the clinicians become highly attractive. In this paper, we proposed a CAD method which realized the automatic detection and annotation of EEC lesions in gastroscopic images. The proposed method initially utilized an advanced Deep Learning (DL) network Deeplabv3+ to obtain a preliminary prediction of EEC regions. Then, a post-processing step which referenced the clinical requirements was designed and applied to get the final annotation results. Totally 3190 gastroscopic images of 732 patients were used in this work. The final experimental results show that the EEC detection rate of our method was 97.07%, and the mean Dice Similarity Coefficient (DSC) was 74.01%, which are higher than those of other state-of-the-are DL-based methods. In addition, the false positive output of our method is fewer. Therefore, the proposed method offers a good potential to aid the clinical diagnoses of EEC.