Deep Convolutional Neural Networks for WCE Abnormality Detection: CNN Architecture, Region Proposal and Transfer Learning.
Libin Lan,Chunxiao Ye,Chengliang Wang,Shangbo Zhou
DOI: https://doi.org/10.1109/access.2019.2901568
IF: 3.9
2019-01-01
IEEE Access
Abstract:Wireless capsule endoscopy (WCE) plays an important role in the diagnosis of gastrointestinal diseases. However, it is very time-consuming and fatiguing for a physician to review a large number of WCE images. Some methods to address this problem have recently been presented. However, these methods generally employ classification algorithms to discriminate abnormal from normal images, which do not localize, recognize, or detect abnormal patterns in abnormal images. We sought to identify a better method for the WCE abnormal pattern detection. In this paper, convolutional neural networks (CNNs) are used to implement detection function, and several methods are also adopted to boost the performance of WCE abnormality detection from aspects of the CNN architecture, region proposal, and transfer learning. First, we present a deep cascade network, namely, CascadeProposal, trained end-to-end to generate a small number of region proposals with high-recall by a region proposal rejection module and to simultaneously detect abnormal patterns using a detection module. Second, we use a multiregional combination (MRC) method to obtain good coverage of the regions of interest and employ the salient region segmentation (SRS) method to capture accurate region locations. Third, we use the dense region fusion (DRF) method for object boundary refinement. Fourth, we introduce negative category (Neg) and transfer learning (TL) strategies into our CNNs to obtain a better model performance. The extensive experiments are performed on our WCE image dataset of more than 7k annotated images. A final mean average precision (mAP) of 70.3% and a better mAP of 72.3% can be achieved via CascadeProposal with ZF and Fast R-CNN with VGG-16 networks, respectively, using MRC+Neg+TL method in the training stage and MRC+DRF+SRS method in the testing stage. The comprehensive results demonstrate that our method is efficient and effective for WCE abnormality detection with high-localization accuracy.