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