论文标题
基于多尺度特征融合的胃肠道息肉和肿瘤检测与WCE序列
Gastrointestinal Polyps and Tumors Detection Based on Multi-scale Feature-fusion with WCE Sequences
论文作者
论文摘要
无线胶囊内窥镜（WCE）已被广泛用于筛查胃肠道（GI）疾病，尤其是小肠，这是由于其非侵入性和无痛的想象的优势，对整个消化道的毫无痛苦的想象。因此，在本文中，我们提出了一个\ textbf {t} wo stage \ textbf {m} ulti-scale \ textbf {f} auture-fusion-fusion学习网络（\ textbf {tmff {tmfnet}），以自动检测小型静脉息肉和tumors in wce图像序列。具体而言，TMFNET由病变检测网络和病变识别网络组成。其中，前者根据传统的更快的R-CNN网络改善了特征提取模块和检测模块，并重新调整了区域提案网络（RPN）模块中锚的参数；后者将残留结构和特征pyramid结构结合在一起，用于构建基于特征型融合的小小的速率，并改善了型号的速率，并将其重新固定，并改善了型号的效率，并将其重新固定，并将其重新固定。实验中的WCE图像，总共有123,092个病变区域用于训练本文的检测框架。在实验中，在医院胃肠病学部提供的实际WCE图像数据集上对检测框架进行了训练和测试。最终模型在RPM上的敏感性，误报和准确性分别为98.81 $ \％$，7.43 $ \％$和92.57 $ \％$。本文提出的算法模型显然优于检测效果和性能中的其他检测算法
Wireless Capsule Endoscopy(WCE) has been widely used for the screening of gastrointestinal(GI) diseases, especially the small intestine, due to its advantages of non-invasive and painless imaging of the entire digestive tract.However, the huge amount of image data captured by WCE makes manual reading a process that requires a huge amount of tasks and can easily lead to missed detection and false detection of lesions.Therefore, In this paper, we propose a \textbf{T}wo-stage \textbf{M}ulti-scale \textbf{F}eature-fusion learning network(\textbf{TMFNet}) to automatically detect small intestinal polyps and tumors in WCE image sequences. Specifically, TMFNet consists of lesion detection network and lesion identification network. Among them, the former improves the feature extraction module and detection module based on the traditional Faster R-CNN network, and readjusts the parameters of the anchor in the region proposal network(RPN) module;the latter combines residual structure and feature pyramid structure are used to build a small intestinal lesion recognition network based on feature fusion, for reducing the false positive rate of the former and improve the overall accuracy.We used 22,335 WCE images in the experiment, with a total of 123,092 lesion regions used to train the detection framework of this paper. In the experiment, the detection framework is trained and tested on the real WCE image dataset provided by the hospital gastroenterology department. The sensitivity, false positive and accuracy of the final model on the RPM are 98.81$\%$, 7.43$\%$ and 92.57$\%$, respectively.Meanwhile,the corresponding results on the lesion images were 98.75$\%$, 5.62$\%$ and 94.39$\%$. The algorithm model proposed in this paper is obviously superior to other detection algorithms in detection effect and performance
