Automatic Polyp Detection from Endoscopy Videos

  • Bilal Taha

Student thesis: Master's Thesis

Abstract

In the last decade, major advancements in the field of Deep Learning (DL) have enabled the development and improvement within the image/signal processing and computer vision applications in diverse fields such as medical imaging, sports and fitness, security as well as business operations. DL can be seen as an extension to Artificial Neural Network (ANN) that incorporate more depth and multipurpose layers which exploit the raw data in different levels of abstraction. The introduction of different level of perception and abstraction have led to many advancements in the field of computer vision and image processing. Moreover, current research has demonstrated the effectiveness of the Convolutional Neural Network (CNN) in the field of image understanding and analysis. However, the dependence of CNN on huge number of annotated samples limit its use in the field of medical imaging. This fact is because of either the costly process for annotating a medical database or the challenges in acquiring sufficient labeled one. Colorectal cancer is one of the major causes of cancer deaths worldwide. To achieve early cancer screening, detecting the presence of polyps in the colon tract is the preferred technique. Manual clinical inspection to detect polyps is the preferred technique to detect polyps. However, many limitations are associated with this technique such as the dependence on the examiner level of expertise and equipment limitations which could result to either false or missed polyps. Therefore, computer aided diagnosis system has been used to help the medical expert and to provide more accurate diagnosis. Since the computer aided diagnosis systems have been introduced, many types of algorithms have been proposed in the literature utilizing different types of features and classifiers. This thesis proposes to employ a CNN as a Transfer Learning (TL) scheme. The proposed approach relies on pre-trained architectures which have been trained on colossal natural images (ImageNet). Contrary to the conventional methods which either perform fine-tuning or train the CNN from scratch, we utilize the CNN output features as an input to train a machineflearning classifier such as Support Vector Machine (SVM) and SoftMax. The polyp samples are first divided into patches where then the learned weights from the pre-trained nets are used to extract deep features from the patches to be used then for the decision making process. Moreover, batch based and data augmentation approaches have also served as a database enlargement techniques were enough samples for training and testing is obtained. Furthermore, a tracking algorithm have been proposed to employ both intensity and color information from the polyp template. The problem is modeled using affine transformation with inverse compositional algorithm using steepest descent as an optimization scheme. The efficiency of the presented framework is demonstrated on public database named CVC ColonDB, in which the experimental results indicate that our methodology significantly out performs other competitive paradigms. Furthermore, the TL scheme developed have been tested on other medical images namely single cell microscopic pap-smear images. This type of images is used for the early detection and classification of Cervical cancer. The results obtained with the second type of cancer confirm the efficiency of this approach for the detection and classification purposes. Indexing Terms: Automatic Polyp Detection, Deep Learning, Transfer Learning, Endoscopy, Tracking, Colorectal Cancer, Polyps, Cervical Cancer. ii
Date of AwardDec 2017
Original languageAmerican English
SupervisorNaoufel Werghi (Supervisor)

Keywords

  • Automatic Polyp Detection
  • Deep Learning
  • Transfer Learning
  • Endoscopy
  • Tracking
  • Colorectal Cancer
  • Polyps
  • Cervical Cancer.

Cite this

'