Time series analysis for medical videos

In this project, we aim to design and develop a system for analysing videos from a camera pill. The pill is swallowed and records video of the digestive system - the goal is to be able to automatically detect cancer in the colon. The main idea is to go beyond image based methods exploiting also the time factor of the data. Algorithms to be explored are for example Recurrent Neural Networks and Long Short Term Memory Networks.
Master

Colon cancer is the third most common cause of cancer mortality for both men and women, and it is a condition where early detection is of clear value for the ultimate survival. As statistics show that 15% of male and female above 50 years are at risk, the procedure is recommended on a regular basis (3-5 years) for the population over 50, and from an earlier age for high-risk groups. Colonoscopy, though, is a demanding procedure requiring significant amount of time by specialized physicians, in addition to the discomfort and risks inherent in the procedure. Thus, traditional methods based on colonoscopy are not cost- effective for population-based screening purposes. These features make the method suboptimal for population based screening purposes, so only about 2-3% of the target population is reached at present. Moreover, the cost of a population screening program is prohibitive. In the US, the colonoscopy is the most expensive cancer screening process with annual costs of $10 billion dollars ($1100/person) [3]. In Norway, we have similar costs of around $1000, with a time consumption of about 1 medical-doctor-hour and 2 nurse-hours, per examination. By researching an automatic system for a camera pill (first picture), the aim is to greatly increase the number of patients that can be examined, i.e., making the public health care system more scalable and cost effective, while at the same time reducing the need for intrusive procedures like “bottom-up” examinations like colonoscopy (second picture).

Goal

There are many different problem areas, and some examples include:

  • Detection of different irregularities in the digestive system, like a colon polyp, Chron's disease, Colorectal cancer, etc. using video object tracking, object detection, machine learning or other relevant tools or mechanisms.
  • Scaling such a system to enable population-wide yearly screening
  • Big-data capture and management
  • Efficient processing of video data possibly using both CPUs and GPUs.
  • Research and build a live system for live examinations at the Hospital

Learning outcome

  • Insight into advanced techniques of machine learning
  • Working on a real world application
  • Collaboration with researchers in the topic of machine learning, specifically deep learning
  • Possibility to implement and research a novel approach

Qualifications

  • Programming
  • Motivation
  • Mathematics

Supervisors

  • Håkon Kvale Stensland
  • Pål Halvorsen
  • Michael Riegler
  • Hugo Lewi Hammer, OsloMet, hugoh@oslomet.no

Collaboration partners

OsloMet