Capsule Endoscopy Image Classification with Deep Convolutional Neural Networks

Abstract

Wireless capsule endoscopy (WCE) is quite an advanced, patient-friendly and novel medical equipment for non-invasive gastrointestinal disease detection, which is able to view the entire gastrointestinal tract without pain. Since there are more than 80,000 WCE images for each examination, it usually takes hours for professional clinicians to diagnose all the video data. Therefore, the automatic computer-aided lesion classification technique is highly perspective needed. In this paper, we propose an effective scheme to classify different lesion images acquired by WCE. Firstly, we obtain feature maps of the same resolution by performing a max pooling operation on different convolutional layers, and then quantify the pooled feature maps by trainable weight parameters, and finally one by one convolution kernels are employed to merge the combined quantized feature maps. We enhance the performance of feature extraction by merging multi-level convolutional features, including both low level and high level features. The preliminary experimental result shows that the classification accuracy rate is up to 95.15%, both running speed and recognition rate are higher than traditional machine learning algorithms.