Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Endoscopic capsule spatial awareness by using image feature extraction and machine learning

Athanasiou Sofia

Full record

---

Summary

Wireless Endoscopy Capsule becomes more and more famous since it is the only way for the physicians to have a non-invasive view of the small intestine and identify bleeding, polyps, Crohn’s disease and other abnormalities of the small intestine. Wireless Capsule Endoscopy is happening with the patient swallowing an electronic capsule, the capsule includes a camera, that way is taking photographs of the Gastrointestinal Track as it moves from one organ to another. Those pictures are transferred to a recorder and from there a software is producing a video stream using those frames. The journey of the capsule in gastrointestinal track might last 8 or more hours based on the movement of the track, during that time it will collect more than 50,000 to 100,000 pictures producing a long video. Examining that amount of frames, it is a very time-consuming procedure and a tiring task for gastroenterologists. There are many researches that try to reduce that time using extra processing of the videos. Nowadays, with the evolution of Machine Learning and Artificial Intelligence for image classification, they are a great supporting tool for physicians.This thesis studies the spatial information extraction problem for Wireless Endoscopy Capsule using Machine Learning techniques. Being able to locate the endoscopy capsule can provide doctors with the opportunity to examine only the part of the gastrointestinal track that is of their most interest. Moreover, producers of endoscopy capsules can use such techniques in order to save battery, by reducing the pictures tracked in areas that are not of the most importance.In order to resolve that problem, we developed three different Convolution Neural Networks models, which achieve the automatic classification of the pictures they receive to the corresponding organ of the digestive system (esophagus, stomach, small intestine and large intestine). The data we used to train and verify our models are from the same collection as in the Thesis of Mr. A. Polydorou, at the Electrical and Computer Engineering Department of Technical University of Crete in 2018. For the feature extraction part of the model, Convolutional two-dimensional (Conv2D) and Maxpooling two-dimensional (Max2D) were used with ReLU activation function. For the classification were used Flatten and Dense layers with activation function softmax function. In one model there was also use of Dropout in order to randomly disconnect nodes and their edges.For the validation of the performance of our models, we used metrics such as Accuracy(Acc), Sensitivity(Sens), Specificity(Spe), Error Rate(Err) and Precision(Pre). Since it is a multi-class classification, those values are calculated for each organ (or class) individually. The best performance among our models holistically comes from Model 2 where is providing for esophagus: Accuracy of 95.16%, Error Rate of 4.83%, Precision of 84.84%, Specificity of 94.73% and Sensitivity of 96.55%. For stomach provides, Accuracy of 91.93%, Error Rate of 8.06%, Precision of 90.9%, Specificity of 96.55% and Sensitivity of 81.08%. For small intestine, is performing with Accuracy of 95.96%, Error Rate of 4.03%, Precision of 92.85%, Specificity of 97.89% and Sensitivity of 89.65%. Finally, for large intestine, Accuracy of 99.19%, Error Rate of 0.8%, Precision of 96.66%, Specificity of 98.94% and Sensitivity of 100%.

Available Files

Services