School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, China.
Endoscopy Center, Department of Gastroenterology, Shanghai East Hospital, Tongji University School of Medicine, Shanghai, 200120, China.
J Imaging Inform Med. 2024 Oct;37(5):1-11. doi: 10.1007/s10278-024-01096-9. Epub 2024 Apr 8.
Capsule endoscopy (CE) is non-invasive and painless during gastrointestinal examination. However, capsule endoscopy can increase the workload of image reviewing for clinicians, making it prone to missed and misdiagnosed diagnoses. Current researches primarily concentrated on binary classifiers, multiple classifiers targeting fewer than four abnormality types and detectors within a specific segment of the digestive tract, and segmenters for a single type of anomaly. Due to intra-class variations, the task of creating a unified scheme for detecting multiple gastrointestinal diseases is particularly challenging. A cascade neural network designed in this study, Cascade-EC, can automatically identify and localize four types of gastrointestinal lesions in CE images: angiectasis, bleeding, erosion, and polyp. Cascade-EC consists of EfficientNet for image classification and CA_stm_Retinanet for lesion detection and location. As the first layer of Cascade-EC, the EfficientNet network classifies CE images. CA_stm_Retinanet, as the second layer, performs the target detection and location task on the classified image. CA_stm_Retinanet adopts the general architecture of Retinanet. Its feature extraction module is the CA_stm_Backbone from the stack of CA_stm Block. CA_stm Block adopts the split-transform-merge strategy and introduces the coordinate attention. The dataset in this study is from Shanghai East Hospital, collected by PillCam SB3 and AnKon capsule endoscopes, which contains a total of 7936 images of 317 patients from the years 2017 to 2021. In the testing set, the average precision of Cascade-EC in the multi-lesions classification task was 94.55%, the average recall was 90.60%, and the average F1 score was 92.26%. The mean mAP@ 0.5 of Cascade-EC for detecting the four types of diseases is 85.88%. The experimental results show that compared with a single target detection network, Cascade-EC has better performance and can effectively assist clinicians to classify and detect multiple lesions in CE images.
胶囊内镜（CE）在胃肠道检查过程中具有非侵入性和无痛的特点。然而，胶囊内镜会增加临床医生对图像进行审查的工作量，容易导致漏诊和误诊。目前的研究主要集中在二分类器上，少数研究针对四种以下异常类型的多分类器，以及特定消化道段的探测器和单一类型异常的分段器。由于类内差异，创建用于检测多种胃肠道疾病的统一方案的任务极具挑战性。本研究设计的级联神经网络 Cascade-EC 可以自动识别和定位 CE 图像中的四种胃肠道病变：血管扩张、出血、糜烂和息肉。Cascade-EC 由用于图像分类的 EfficientNet 和用于病变检测和定位的 CA_stm_Retinanet 组成。作为 Cascade-EC 的第一层，EfficientNet 网络对 CE 图像进行分类。作为第二层的 CA_stm_Retinanet，在分类图像上执行目标检测和定位任务。CA_stm_Retinanet 采用 Retinanet 的通用架构。其特征提取模块是来自 CA_stm_Block 堆叠的 CA_stm_Retinanet。CA_stm_Block 采用分-转-合策略，并引入坐标注意力。本研究的数据来自上海东方医院，由 PillCam SB3 和 AnKon 胶囊内镜采集，包含 2017 年至 2021 年间 317 名患者的总共 7936 张图像。在测试集中，Cascade-EC 在多病变分类任务中的平均精度为 94.55%，平均召回率为 90.60%，平均 F1 得分为 92.26%。Cascade-EC 检测四种疾病的平均 mAP@0.5 为 85.88%。实验结果表明，与单一目标检测网络相比，Cascade-EC 具有更好的性能，可以有效帮助临床医生对 CE 图像中的多种病变进行分类和检测。
