The Use of Convolutional Neural Artificial Intelligence Network to Aid the Diagnosis and Classification of Early Esophageal Neoplasia. A Feasibility Study
Refereed conference paper presented and published in conference proceedings


Times Cited
Web of Science1WOS source URL (as at 11/07/2020) Click here for the latest count
Altmetrics Information
.

Other information
AbstractIntroduction
The diagnosis and classification of early esophageal neoplasia by the novice endoscopist is difficult. Recently, the use of artificial intelligence to aid in pattern recognition and image analysis is becoming popular. The aim of the current study is to assess if a convolutional neural artificial intelligence network could aid the diagnosis and classification of early esophageal neoplasia.

Methods
The endoscopic images in patients with normal esophagus or those suffering from early esophageal neoplasia were obtained using narrow band imaging and magnifying endoscopy (GIF-FQ260Z, Olympus Medical, Japan) and (GIF-HQ290 Olympus Medical, Japan). The images were classified according to the Japan Esophageal Society into Type A, B1, B2 and B3 vessels and confirmed by histological assessment on endoscopic or surgical resection. An experienced endoscopic then identified regions of Interest (RoI) in the endoscopic images. The images were then computer processed by randomly sampling 1000 patches of size 50 x 50. A patch is in RoI category if it contains more than 50% RoI pixels, otherwise it is in background category.

Supervised Learning via Convolutional Neural Network.

A supervised learning task takes training patches and their category labels as input, and produces an algorithm that predicts the label for a new patch. For the patch classification task, we use convolutional neural network (CNN), which is proven to be very effective for image classification. The process of training is to modify the network's weights to better align with the training data. The input layer is of dimension 3 x 50 x 50, because the patch is of size 50 x 50 and it has three color channels (red, green and blue). Each 3 x 5 X 5 small window in the input layer produces a single data entry in the second layer. The next layer is due to max-pooling, which summarizes the features to make the classification result insensitive to shift and rotation. The same convolution and pooling method is applied again in the next two layers. The last 5 layers are fully connected. Two entries in the last layer are the output of the neural network.

Results
A total of 218 endoscopic images of normal and neoplastic esophagus were obtained. 218000 patches (80740 RoI patches and 137260 background patches) were generated. Among the 218000 patches, 90% of them are selected to be training set and the rest were used for testing. The network takes 2.5 days to train on a GTX 980 Ti GPU.

The overall diagnostic accuracy was 79.38%. The sensitivity, specificity, positive predictive value, negative predictive value were 73.41%, 83.54%, 72.09%, 84.44% respectively.

Conclusions
The use of convolutional neural artificial intelligence network to aid the diagnosis and classification of early esophageal neoplasia is feasible. The results could be further improved with a larger archive of endoscopic images.
All Author(s) ListZhang CZ, Ma L, Uedo N, Matsuura N, Tam P, Teoh AY
Name of ConferenceDigestive Disease Week 2017
Start Date of Conference07/05/2017
End Date of Conference09/05/2017
Place of ConferenceChicago, Illinois
Country/Region of ConferenceUnited States of America
Proceedings TitleGastrointestinal Endoscopy
Title of PublicationGASTROINTESTINAL ENDOSCOPY
Year2017
Month5
Volume Number85
Issue Number5 Supplement
PublisherElsevier
PagesAB587 - AB588
ISSN0016-5107
LanguagesEnglish-United Kingdom
Web of Science Subject CategoriesGastroenterology & Hepatology;Gastroenterology & Hepatology

Last updated on 2020-12-07 at 04:49