Detection Of Lesions In Wireless Capsule Endoscopy Images Based On Multi-scale Full Convolution Network

With the acceleration of life rhythm and the increase of social pressure,more and more people suffer from gastrointestinal diseases because of irregular diet.Wireless Capsule Endoscopy technology has become the first choice for examining patients’ gastrointestinal tract.A wireless capsule endoscopy check produce about 60,000 images,which brings a heavy burden to the doctors’ screening work.How to quickly and effectively recognize and detect pathological changes in the 60,000 images is an urgent problem to be solved.Deep neural network has unique advantages in image processing,and medical image analysis based on deep learning provides an opportunity to solve this problem.Wireless capsule endoscopy image is very different from the traditional scene image.It has the characteristics of small data set,low pixel resolution,diverse lesion size,fuzzy detail features,and more interference of digestive tract impurities,which makes the traditional machine learning technology unable to meet the requirements of wireless capsule endoscope image recognition and detection(such as detection of gastrointestinal polyps,ulcers,bleeding points).Aiming at the problem,by introducing the idea of skip connections in residual network and multi-scale convolution kernel in Inception network,a multi-scale convolution with shortcut connection structure is constructed,which can preserve the shallow features of images and combine them with the high-level features.At the same time,multi-scale convolution kernel is used to enhance the effective extraction of different scale lesion features.The full convolution neural network composed of this module has fewer free parameters than the traditional convolution neural network,which is more suitable for the characteristics of wireless capsule endoscopy image and the needs of disease detection.From the experimental results,the neural network model can detect lesions on 40000 wireless capsule endoscopy images,and the accuracy,sensitivity and specificity of the neural network model are 97.84%,98.05% and 97.67%,respectively.The neural network model is superior to the classical depth residual network Res Net and classic multi-scale Inception-v4 model in convergence and scale tolerance.