Abstract
BACKGROUND: Polypoid lesions of the gallbladder (PLGs) are common lesions that can be classified as either nonneoplastic or neoplastic polyps. Surgical resection is recommended for neoplastic polyps, and accurate preoperative identification of neoplastic polyps is needed to guide appropriate management. However, accurately distinguishing neoplastic polyps remains challenging. Therefore, the aim of this study is to establish a preoperative prediction model for neoplastic polyps on the basis of a convolutional neural network (CNN) model using ultrasound images and evaluate its reliability. METHODS: This was a multicentre retrospective study. All included cases were divided into a training set, an internal test set, and an external test set. A CNN model was established using the Inception-V3 model, and the ultrasound images from the training set were input into the CNN for feature processing. The internal and external test set images were subsequently used to assess the predictive performance of the CNN model, which was then compared with the diagnostic performance of three sonographers with different levels of experience and an ultrasound feature-based nomogram model. RESULTS: A total of 380 cases (921 images in total) were retrospectively collected, with 194 cases in the training set (547 images in total), 83 cases in the internal test set (234 images in total), and 103 cases in the external test set (140 images in total). The areas under the curves (AUCs) of the CNN model were 0.896 and 0.852 in the internal and external test sets, respectively. In addition, the CNN model outperformed the three sonographers with varying levels of experience (AUC =0.687, 0.703, and 0.803, respectively), but was comparable to the nomogram model (AUC =0.880) in terms of diagnostic efficacy. CONCLUSIONS: The CNN model, which is based on ultrasound images, has demonstrated relatively good predictive performance in preoperatively identifying neoplastic polyps and is highly important for assisting in the selection of treatment methods for PLG patients.