Color-guided deformable convolution network for intestinal metaplasia severity classification using endoscopic images

Abstract

Abstract Objective. Intestinal metaplasia (IM) is a common precancerous condition for gastric cancer, and the risk of developing gastric cancer increases as IM worsens. However, current deep learning-based methods cannot effectively model the complex geometric structure of IM lesions. To accurately diagnose the severity of IM and prevent the occurrence of gastric cancer, we revisit the deformable convolution network (DCN), propose a novel offset generation method based on color features to guide deformable convolution, named color-guided deformable convolutional network (CDCN). Approach. Specifically, we propose a combined strategy of conventional and deep learning methods for IM lesion areas localization and generating offsets. Under the guidance of offsets, the sample locations of convolutional neural network adaptively adjust to extract discriminate features in an irregular way that conforms to the IM shape. Main results. To verify the effectiveness of CDCN, comprehensive experiments are conducted on the self-constructed IM severity dataset. The experimental results show that CDCN outperforms many existing methods and the accuracy has been improved by 5.39% compared to DCN, reaching 84.17%. Significance. To the best of our knowledge, CDCN is the first method to grade the IM severity using endoscopic images, which can significantly enhance the clinical application of endoscopy, achieving more precise diagnoses.