Md Jahin Alam1, Rifat Bin Rashid1, Shaikh Anowarul Fattah1, Mohammad Saquib2. 1. Department of Electrical and Electronic EngineeringBangladesh University of Engineering and Technology Dhaka 1000 Bangladesh. 2. Department of Electrical EngineeringThe University of Texas at Dallas Richardson TX 75080 USA.
Abstract
Background: The emergence of wireless capsule endoscopy (WCE) has presented a viable non-invasive mean of identifying gastrointestinal diseases in the field of clinical gastroenterology. However, to overcome its extended time of manual inspection, a computer aided automatic detection system is getting vast popularity. In this case, major challenges are low resolution and lack of regional context in images extracted from WCE videos. Methods: For tackling these challenges, in this paper a convolution neural network (CNN) based architecture, namely RAt-CapsNet, is proposed that reliably employs regional information and attention mechanism to classify abnormalities from WCE video data. The proposed RAt-CapsNet consists of two major pipelines: Compression Pipeline and Regional Correlative Pipeline. In the compression pipeline, an encoder module is designed using a Volumetric Attention Mechanism which provides 3D enhancement to feature maps using spatial domain condensation as well as channel-wise filtering for preserving relevant structural information of images. On the other hand, the regional correlative pipeline consists of Pyramid Feature Extractor which operates on image driven feature vectors to generalize and propagate local relationships of pixels from WCE abnormalities with respect to the normal healthy surrounding. The feature vectors generated by the pipelines are then accumulated to formulate a classification standpoint. Results: Promising computational accuracy of mean 98.51% in binary class and over 95.65% in multi-class are obtained through extensive experimentation on a highly unbalanced public dataset with over 47 thousand labelled. Conclusion: This outcome in turn supports the efficacy of the proposed methodology as a noteworthy WCE abnormality detection as well as diagnostic system.
Background: The emergence of wireless capsule endoscopy (WCE) has presented a viable non-invasive mean of identifying gastrointestinal diseases in the field of clinical gastroenterology. However, to overcome its extended time of manual inspection, a computer aided automatic detection system is getting vast popularity. In this case, major challenges are low resolution and lack of regional context in images extracted from WCE videos. Methods: For tackling these challenges, in this paper a convolution neural network (CNN) based architecture, namely RAt-CapsNet, is proposed that reliably employs regional information and attention mechanism to classify abnormalities from WCE video data. The proposed RAt-CapsNet consists of two major pipelines: Compression Pipeline and Regional Correlative Pipeline. In the compression pipeline, an encoder module is designed using a Volumetric Attention Mechanism which provides 3D enhancement to feature maps using spatial domain condensation as well as channel-wise filtering for preserving relevant structural information of images. On the other hand, the regional correlative pipeline consists of Pyramid Feature Extractor which operates on image driven feature vectors to generalize and propagate local relationships of pixels from WCE abnormalities with respect to the normal healthy surrounding. The feature vectors generated by the pipelines are then accumulated to formulate a classification standpoint. Results: Promising computational accuracy of mean 98.51% in binary class and over 95.65% in multi-class are obtained through extensive experimentation on a highly unbalanced public dataset with over 47 thousand labelled. Conclusion: This outcome in turn supports the efficacy of the proposed methodology as a noteworthy WCE abnormality detection as well as diagnostic system.
Entities:
Keywords:
GI tract; Wireless capsule endoscopy; attention mechanism; deep CNN; pyramid
Authors: Debesh Jha; Sharib Ali; Steven Hicks; Vajira Thambawita; Hanna Borgli; Pia H Smedsrud; Thomas de Lange; Konstantin Pogorelov; Xiaowei Wang; Philipp Harzig; Minh-Triet Tran; Wenhua Meng; Trung-Hieu Hoang; Danielle Dias; Tobey H Ko; Taruna Agrawal; Olga Ostroukhova; Zeshan Khan; Muhammad Atif Tahir; Yang Liu; Yuan Chang; Mathias Kirkerød; Dag Johansen; Mathias Lux; Håvard D Johansen; Michael A Riegler; Pål Halvorsen Journal: Med Image Anal Date: 2021-02-19 Impact factor: 8.545
Authors: Pia H Smedsrud; Vajira Thambawita; Steven A Hicks; Debesh Jha; Thomas de Lange; Michael A Riegler; Pål Halvorsen; Henrik Gjestang; Oda Olsen Nedrejord; Espen Næss; Hanna Borgli; Tor Jan Derek Berstad; Sigrun L Eskeland; Mathias Lux; Håvard Espeland; Andreas Petlund; Duc Tien Dang Nguyen; Enrique Garcia-Ceja; Dag Johansen; Peter T Schmidt; Ervin Toth; Hugo L Hammer Journal: Sci Data Date: 2021-05-27 Impact factor: 6.444