DeepCPC, deep learning model for colorectal polyps classification

Other Title(s)

DeepCPC : نموذج التعلم العميق لتصنيف سلائل أورام القولون و المستقيم

Dissertant

Taha, Dima Hussein Fakhri

Thesis advisor

al-Zubi, Ahmad Ghazi

University

Middle East University

Faculty

Faculty of Information Technology

Department

Computer Science Department

University Country

Jordan

Degree

Master

Degree Date

2020

English Abstract

Colorectal cancer is a silent disease that attacks without warning, but in many cases, treatment is possible when discovered early.

Screening tests play an important role in identifying polyps before they become cancerous, where colonoscopy is more effective compared to other tests.

Over the past few decades, the computer-aided colorectal polyp in colonoscopy has been the subject of research and achieved significant advances.

However, the automatic polyp classification in real-time is still a challenging problem due to utilizing the hand-crafted methods that do not provide discriminating image features.

The advanced deep convolutional neural networks (DCNN) have shown a significant revolution that positively influenced many fields including image classification.

In the domain of colonoscopy images, many limitations could affect the DCNN-based polyp’s classification especially the lack of sufficient amount of patients' training samples, inadequate training time, and needed resources for neural networks.

The work in this thesis aims to develop a deep learning model for classifying colorectal polyps (referred to as DeepCPC), based on discriminative features extracted from deep conventional neural networks.

Specifically, some CNN models pretrained has been used on general-purpose images to apply a transfer learning scheme in the polyp’s classification system.

This is achieved by concatenating a set of discriminating image features extracted from the activations of convolutional layers, then improved feature representations by finetuning a proposed CNN architecture on polyps images through a complete end-to-end training procedure.

The proposed model consists of three main components: lower fine-tuned layers, concatenated image vector, and fully-connected top layers.

The CVCClinicDB dataset is used to evaluate the Deep CPC model, but further patch extraction and image augmentation strategies are applied to enrich the training procedure with more sufficient polyp’s samples.

The experimental results show that the proposed CNN model can achieve an accuracy of 98.4%, which emphasize its efficiency for helping endoscopic physicians to classify polyps and decrease the colorectal polyp miss rate.

Main Topic

Diseases
Information Technology and Computer Science

Topics

No. of Pages

69

Table of Contents

Table of contents.

Abstract.

Abstract in Arabic.

Chapter One : Introduction.

Chapter Two : Background and literature review.

Chapter Three : Methodology and DeepCPC model.

Chapter Four : Implementation and results.

Chapter Five : Conclusion and future work.

References.

American Psychological Association (APA)

Taha, Dima Hussein Fakhri. (2020). DeepCPC, deep learning model for colorectal polyps classification. (Master's theses Theses and Dissertations Master). Middle East University, Jordan
https://search.emarefa.net/detail/BIM-970871

Modern Language Association (MLA)

Taha, Dima Hussein Fakhri. DeepCPC, deep learning model for colorectal polyps classification. (Master's theses Theses and Dissertations Master). Middle East University. (2020).
https://search.emarefa.net/detail/BIM-970871

American Medical Association (AMA)

Taha, Dima Hussein Fakhri. (2020). DeepCPC, deep learning model for colorectal polyps classification. (Master's theses Theses and Dissertations Master). Middle East University, Jordan
https://search.emarefa.net/detail/BIM-970871

Language

English

Data Type

Arab Theses

Record ID

BIM-970871