Endoscopy-assisted lightweight diagnosis system based on transformers for colon polyp detection

作     者：FAN Weiming YU Jiahui JU Zhaojie 

作者机构：School of Automation and Electrical Engineering， Shenyang Ligong University Department of Biomedical Engineering， Zhejiang University School of computing， University of Portsmouth 

出 版 物：《Optoelectronics Letters》 (光电子快报(英文))

年 卷 期：2024年

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 1002[医学-临床医学] 08[工学] 081104[工学-模式识别与智能系统] 080203[工学-机械设计及理论] 100201[医学-内科学(含：心血管病、血液病、呼吸系病、消化系病、内分泌与代谢病、肾病、风湿病、传染病)] 0802[工学-机械工程] 0835[工学-软件工程] 0811[工学-控制科学与工程] 10[医学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by the Zhejiang Provincial Natural Science Foundation of China (No. LQ23F030001) Hangzhou Innovation Team (No. TD2022011) the AiBle project co-financed by the European Regional Development Fund and National Natural Science Foundation of China (52075530) 

摘      要：The integration of endoscopy has significantly propelled the diagnosis and treatment of gastrointestinal diseases， with colonoscopy establishing itself as the primary method for early diagnosis and preventive care in colorectal cancer. Although deep learning holds promise in mitigating missed polyp rates， modern endoscopy examinations pose additional challenges， such as image blurring and atomizing. This study explores lightweight yet powerful attention mechanisms， introducing the Spatial-Channel Transformer—an innovative approach that leverages spatial channel relationships for attention weight calculation. The method utilizes rotation operations for inter-dimensional dependencies， followed by residual transformation， encoding inter-channel and spatial information with minimal computational overhead. Extensive experiments on the CVC-ClinicDB polyp detection dataset， addressing endoscopy pitfalls， underscore the superiority of our Spatial-Channel Transformer over other state-of-the-art methods. The proposed model maintains high performance， even in challenging scenarios.