Deep Convolutional Neural Networks for WCE Abnormality Detection: CNN Architecture, Region Proposal and Transfer Learning

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. INDEX TERMS Convolutional neural networks, medical image analysis, region proposal, transfer learning, wireless capsule endoscopy, WCE abnormality detection.

Review Research of Medical Image Analysis Using Deep Learning

“…The results showed that according to tumor motion and projection angles which exhibits that the CNN based method was more robust and accurate in real-time tumor localization [48]. Lan et al (2019) explored that multiregional combination such as selective search, edge boxes, and abjectness is used to improve object localization that account as essential of the non-rigid and amorphous characteristics to improve object localization [41]. Urban et al (2018) showed ADR aim of colonoscopy and accuracy according of colonoscopies for ADR.…”

“…It is a type of advanced optical fluorescence technology which undergoing application assessments in brain tumor surgery while most of the images distorted and interpreted as non-diagnostic images [40]. In gastrointestinal diseases, new medical imaging technique innovated which known as wireless capsule endoscopy (WCE) to record WCE frame images to detect abnormal patterns [41]. It uses to diagnose of gastrointestinal diseases through a sensor which is quite small to swallow and capture every scenes of anatomical parts that pass through them [41].…”

“…In gastrointestinal diseases, new medical imaging technique innovated which known as wireless capsule endoscopy (WCE) to record WCE frame images to detect abnormal patterns [41]. It uses to diagnose of gastrointestinal diseases through a sensor which is quite small to swallow and capture every scenes of anatomical parts that pass through them [41]. Dermoscopic image is another useful modality and is dermoscopic images that use to skin lesion [42], [43].…”

“…In neurodegenerative diseases, dementia for instance which causes of lessens in memory, language, and lack of wise [22]. That can boost the performance of CNN and improve detection and localization accuracy [41]. For super-pixel image analysis, different structure detection required.…”

Multi‐scale analysis of ulcer disease detection from WCE images

“…Thus, though CNN has great potential for the detection of mucosal breaks in WCE images, in some cases, the handcrafted machine learning approaches may perform better than CNN algorithms. To address the problem of small WCE dataset, Lan et al [31] proposed a deep cascade network architecture (CascadeProposal) pre‐trained via transfer learning (fine‐tuning) only for bleeding, polyp and tumours recognition from WCE images. Ulcer detection from WCE images is still a challenging task.…”