Gastrointestinal (GI) cancers, mainly including esophageal cancer, gastric cancer, and colorectal cancer, are the leading cause of cancer-related deaths worldwide. Most GI cancers go through the stages of precancerous lesions, which can be defined as common conditions associated with a higher risk of developing cancers over time. Early identification of precancerous lesions has been shown to minimize the incidence of GI cancers and substantiate the vital role of screening endoscopy. However, unlike GI cancers, precancerous lesions in the GI tract can be subtle and difficult to detect. Previous studies have demonstrated that convolutional neural network (CNN) trained with endoscopic images can help endoscopists identify GI cancers accurately. Currently, the CNN is the most representative deep learning approach. Thus, this study carries out a deep investigation on the development of deep learning-based methods to diagnose precancerous lesions in the GI tract. Firstly, an esophageal lesion attention network (ELANet) is developed to assist endoscopists in the diagnosis and treatment of multiple types of esophageal lesions, including Barrett esophagus, reflux esophagitis, and esophageal cancer. Two attention units are combined, the spatial attention unit and the channel attention unit, using bilinear strategy to build the ELANet, which can emulate the endoscopists’ diagnosis that focuses on ‘where’ is a lesion-related region, and ‘what’ is informative when analyzing the endoscopic image. Moreover, the ELANet can aggregate with different backbone networks and be trained in an end-to-end manner to make better performance. Then, an intelligent diagnostic (ID) system is constructed to provide an objective assistance in the diagnosis of gastric intestinal metaplasia, a precancerous lesion of gastric cancer, based on a limited number of narrow-band and magnifying iii narrow-band images. Three binary classifiers using the same modified CNNs are independently trained and combined to improve the accuracy and robustness of the ID system. Thirdly, a novel multi-feature fusion system (MFFS) is proposed to assist endoscopists in the detection and histological prediction of colorectal polyps, especially for the adenomatous polyps, which are considered precancerous lesions associated with colorectal cancer. Inspired by the clinical knowledge of polyp diagnosis, the texture features, color features, and features learned by CNNs are fused to build the powerful MFFS. The endoscopic images collected in the above studies are mainly from the Kiang Wu Hospital, Macao and Xiangyang Central Hospital, P.R. China, but the numbers are small. Therefore, transfer learning, data augmentation, bilinear strategy, attention mechanism, and multi-feature fusion are used in combination to overcome the issue of insufficient data. Experimental results show that all of the proposed methods exhibit promising diagnostic performance, which indicates that the proposed methods have great potential to assist endoscopists in the decision-making process. Compared with previous algorithms, the proposed algorithms can achieve good diagnostic results without a large amount of data.