| Wireless Capsule Endoscopy(WCE)is an important gastrointestinal screening technique for early and extensive screening of diseases such as vascular lesion,inflammatory,polyps,ulcers and even cancer in the gastrointestinal tract.WCE produces tens of thousands of images during one examination.It is time-consuming and inefficient to select a small number of key lesion images from a vast,complex set of images.Advances in artificial intelligence technology offer a poosible solution to this problem.However,due to the diverse morphology of gastrointestinal lesions an d the large number of images,current WCE images lesion recognition methods can hardly meet the requirements for both detection accuracy and efficiency.To solve the abovementioned problem,a deep learning-based method for rapid recognition of lesion in WCE images is proposed and software for WCE video recognition is developed.The effectiveness of the proposed method is verified through experiments,which has important theoretical value and practical significance.The main contributions of the present wo rk are as follows:(1)According to the characteristics of WCE images,a high-resolution lightweight convolutional neural network(LHNet)is designed to address the difficulties in WCE images lesion recognition.LHNet consists of four modules,including a lightweight,high-resolution backbone network(LHB),a densely connected residual group(DCRG),a parallel pooling head(PPH)and a global attention branch(GAB).LHB and DCRG have strong feature representation capability and can effectively solve the diff iculty in feature extraction caused by the small size and blurred boundary of gastrointestinal lesions.PPH and GAB can highlight key features,suppress irrelevant information and enhance the feature selection ability of the network,thus improving the classification ability of LHNet for images with high inter-class similarity/intra-class difference.LHNet is mainly constructed by deep separable convolution,based on which a lesion recognition method for WCE images is proposed so that it meets the requirement of lesion recognition accuracy as well as computing efficiency.(2)A comprehensive evaluation of our proposed method is performed using two public datasets with nearly 3,000 WCE images.Results show that the proposed method obtains an overall accuracy of 95.42% and 91.42% on the two datasets,respectively.Compared with two state-of-the art methods,the parameter amount of LHNet is reduced by 90.15% and 78.72%,and the number of floating-point operations by 91.94%and 90.23%,respectively.Therefore,our proposed method outperforms current methods in lesion classification accuracy and computing efficiency.To test the generalisation of the proposed LHNet,an experimental evaluation is conducted using WCE videos of the whole gastrointestinal tract of 16 patients collected from hospital.Results show that abnormal frames containing polyps,erosions and bleeding lesions can be quickly identified,and the test generalization error of the network is only 7.3%.(3)A software for lesion recognition in WCE video is developed using C# and Microsoft Visual Studio platform,which is registered with copyright.Tests show that the software achieves the function of lesion recognition,and runs stably with a computation time of about 5 ms per frame on a PC with ordinar y configuration,which meets the requirement for clinical auxiliary diagnosis. |
PDF Full Text Request