Novel Image Processing Algorithms for Automatic Detection of Colon Cancer

The goal of this PhD project is to develop a program consisting of algorithms to automate colonoscopy screening procedures. The algorithms will analyze video-recordings from colonoscopies (eventually from wireless capsule endoscopy, WCE). The developed algorithms should: 1) automatically identify and tag suspicious lesions, especially precancerous polyps, in videos of the human colon obtained by a standard colonoscopy or a WCE 2) automatically identify patients at high risk of having or developing colorectal cancer (CRC) 3) tolerate lesion variability in order to detect all types of lesions 4) help clinicians reduce the lesion miss rate during colonoscopy examination The algorithms are based on deep learning technologies that have recently been applied to successfully analyze large natural image based datasets. The program will highlight areas of the images which are abnormal in appearance and which may be pathological and thus deserve closer examination by a medical doctor. The algorithms will translate a doctor's experience and expertise in recognizing and identifying pathological states. Trying to re-create this with a software program entails many complicated steps such as: identifying the defining features of different types of abnormalities; highlighting how these features are unique compared to the healthy surrounding tissue; correcting for the inevitable movement in the video footage, either from involuntary body movements or from the movement of the camera itself thorough the body; and lastly, writing the algorithms/software which will interpret and identify those aspects of a suspected abnormality accurately enough to make a dependable reading of the video footage possible. The focus is on the automatic detection of polyps in the colon which are the main cause of CRC. Different algorithms are developed to overcome the challenges associated with the complexity of the human colon. For example, the models are developed to understand the complex environment of the inner lining of the colon (mucosa) and to distinguish various polyp-like structures which can mimic real polyps. The only challenge that remains unsolved is the lack of availability of a large labelled dataset of lesion images and videos which is essential for the development of an efficient model that can detect all kinds of polyps. Collecting medical data is difficult because 1) it is ethically sensitive information, and 2) it is not easy for computer scientists to understand medical data; i.e., clinicians must interpret and label the data.

Knowledge was developed about various computer-aided diagnostic algorithms based on deep learning technologies to automatically detect tumors in different medical imaging modalities, especially colon polyps in colonoscopy and wireless capsule endoscopy (WCE) with the result that endoscopies can be performed more efficiently and accurately, reducing overall time spent and the polyp-miss rate. Advanced technical skills were developed which will have significant impacts on the development and enhancement of algorithms for medical data analysis, thereby improving healthcare and saving lives. Plans are to implement the developed automatic polyp detection model on cloud and use 5G infrastructure to interconnect the model with WCE and develop an optimal colon assessment system for screening the population that would be cost efficient and widely available. The results achieved show that engineers and clinicians can collaborate to solve technologically challenging problems faced by clinicians.

This project is about designing and implementing advanced image processing algorithms to analyse the video footage taken of a person's gastrointestinal (GI) tract. These image processing algorithms will constitute a unique, innovative & extremely complicated software programme to be used in conjunction with a pill camera. There are a number of challenges associated with this project; for example, we presently do not have a camera & sensor which can be introduced into the human body, so we will need to create an artificial lab environment in which we can capture accurate video footage for study. Another challenge is that we need to identify sensor technologies which are capable of capturing the required amount of detail in the video footage. If the camera cannot see clearly, and the images are not clear, then it will be next to impossible to write accurate algorithms that will distinguish pathological areas from healthy areas. Another challenge is that this software will have to translate a doctor's experience & expertise in recognizing & identifying pathological states in the colon. The challenge is to properly identify the abnormalities on the image itself, to identify an aspect of that abnormality that can be measured in some way, & to then write algorithms which interpret & identify those aspects of a suspected abnormality accurately enough to make a dependable reading of the video footage possible. The long-term goal is to be able to make an automatic diagnosis by swallowing a pill camera and using the diagnostic software.