消化道胶囊内窥镜智能控制及病变识别的数学模型与算法

Capsule endoscopy (CE) has been an innovative medical technology in recent years and the entire digestive tract examination can be comfortably performed with the application of CE. However, the traditional painful and high-risk gastrointestinal endoscopy cannot be replaced by CE at present. The reasons are as follows: firstly, each case of CE producing more than 50,000 pictures has greatly increased the workload of the reading doctor. Secondly, CE cannot provide enough coverage and accurate position information for certain parts of the digestive tract, which may give rise to missed lesions and the misjudgment of anatomical regions. In order to solve these problems, our project intends to rely on the large sample inspection image data set of the high-level clinical center and is based on the three-dimensional vision solution accumulated in the early stage of the research group, the dual-drive technology of model data and the intelligent recognition technology of lesion images. With the above foundations and supports, the coverage rate of digestive tract examinations would be improved through proposing innovative mathematical models, and the image reading efficiency improved by innovative algorithms would reduce invalid and redundant information during examinations. Moreover, intelligent identification of lesions and accurate positioning of the sited of digestive tract anatomy would be realized through innovative intelligent lesion and anatomical recognition technology. By integrating the above innovative technologies together, a new smart CE innovative application mode was built to realize the convenient intelligent self-examination of the gastrointestinal tract at home, which may provide innovative solutions for early diagnosis and control of major gastrointestinal diseases.

胶囊内窥镜是近年问世的一项可舒适完成消化道全程检查的创新医疗技术，如能解决以下技术难题未来有望取代传统的痛苦而高风险的胃肠镜检查：胶囊内窥镜每例检查产生超过5万张的巨量图片极大增加了阅片医生的工作量；胶囊内窥镜无法对消化道某些部位提供足够的覆盖及精确的位置信息，因而出现病变的漏检及解剖部位的误判。为解决以上问题，本课题拟依托高水平临床中心的大样本检查图片数据集，基于课题组前期积累的三维视觉解决方案、模型数据双驱动技术及病灶图像智能识别技术的基础上，通过数学方法提出创新数学模型以提升消化道检查覆盖率，通过创新算法减少检查中的无效与冗余信息以提升阅片效率，通过创新智能病灶与解剖部位识别技术以实现对病灶的智能识别与消化道解剖部位的准确定位，并整合以上创新技术构建新型智能胶囊内窥镜创新应用模式，实现在家中即可对胃肠道进行便捷的内窥镜影像智能自检，为胃肠重大疾病早诊防控提供创新解决方案。

胶囊内窥镜是一项可舒适完成消化道全程检查的创新医疗技术，但该检查产生的海量图片对阅片医生增加的工作负担，以及无法覆盖消化道特定部位和提供精确的位置信息极大地限制了它的应用。为了解决这些问题，本项目拟整合数学模型、创新算法和智能识别技术创建胶囊内窥镜的新型应用模式，基于此开展的研究产出了以下成果：1、建立了一个分布式、大容量、可存储巨量图片的数据库平台EndoNet，为图像识别算法的研发提供便利；2、提出基于状态转移概率建模的寻优算法和基于深度强化学习的算法来优化胶囊内窥镜的体位选择，并建立了七种内窥镜图片的三维建模/智能识别算法；3、基于时序Transformer的胃部部位识别、基于多尺度Transformer的细粒度多病灶检测、基于帧重要性辅助稀疏子集选择的关键帧提取和多层次域适应的无线胶囊内镜图像超分辨率，这四大特征实现了内窥镜图像的智能定位和智能阅片；4、开发了“GICE智能胶囊检查APP”、“GICE智能图像分析软件”和“康华云镜阅片中心”云阅片平台，实现智能检查、AI阅片及异地云端阅片。取得的这些研究成果为实现家庭内对胃肠道进行内窥镜影像智能便捷筛查提供了可能性，为胃肠疾病的早诊断早治疗提供了新型解决方案。