Abstract

Info

Title:: TE-Deformable DETR based detection algorithm for esophageal squamous cell carcinoma

Author(s):: JIANG Chuandi1; 2; ZHANG Jiatian1; 2; LIANG Yan3; FENG Yadong3; DANG Shijie2; ZHAO Lingxiao2; 1. Division of Life Sciences and Medicine, Biomedical Engineering College (Suzhou), University of Science and Technology of China, Hefei 230026, China 2. Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215162, China 3. Department of Gastroenterology, Zhongda Hospital, Southeast University, Nanjing 210009, China

Keywords:: Keywords: esophageal squamous cell carcinoma object detection texture enhancement search attention

Abstract:: Abstract: To address the issues of visual similarity in color and texture between esophageal squamous cell carcinoma (ESCC) lesions and healthy regions in digestive endoscopy, and the limited accuracy of convolutional neural network based object detection models, a texture-enhanced deformable DETR (TE-Deformable DETR) based detection algorithm is developed. Firstly, the pre-training strategy of UP-DETR is used to pre-train 110 079 images from a public gastrointestinal dataset, thereby enhancing the generalization ability of the Transformer. Next, a preprocessing algorithm for texture and color enhancement is proposed to effectively enhance the contrast between lesions and the background. Then, TE-Deformable DETR is designed, which elevates the models sensitivity to subtle texture changes through a multi-scale texture feature enhancement projection layer, and further enhances the models capacity for lesion feature capture by integrating a search attention mechanism. The proposed method exhibits higher detection accuracy on a multi-center ESCC dataset. Experimental results show that compared with the baseline Deformable DETR algorithm, the improved algorithm increases mAP and mAP50 by 5.1% and 6.1%, respectively, with the recall rate reaching 98.7%.