Cancer & Cardiac Imaging

Oncological Image Analysis

The main objective of this project is to develop image analysis techniques in order to help clinicians assess the oncological imaging data for better diagnosis and treatment planning. In particular, we are considering colorectal cancer (CRC) imaging data as a case study. The key clinical questions in the study of CRC imaging are:

  1. What is the margin of clearance (circumferential resection margin, from now on referred to as CRM) of the tumour affected region from its neighbouring anatomical landmarks?
  2. Are the lymph nodes surrounding the colon/rectum are affected?

Image analysis challenges involved in the assessment of CRC data include: poor contrast, imaging artefacts, variety and scarcity of features along the boundary regions, and small size of regions of interests (ROIs)especially of the lymph nodes. So far in this project, we have proposed two partialdifferential equations (PDE) based contour evolution methods, also known as level set methods, to obtain the segmentation of the CRC images. This enables us to estimate the CRM quantitatively and in a 3-dimensional manner that is not easily achievable manually.
We have also developed a method to estimate distribution of intensity values in an ROI of the image especially when a small number of observed pixels are available. This method has shown some promising results in assessing the status of lymph nodes. The second major objective of this project is to compare continuous spatial regularisation techniques, such as level sets, with discrete ones, such as graph cuts. We have begun by extending our work on level sets, and, since MSR have considerable expertise in graph cuts, we have explored spectral graph theory. These techniques, which have been proposed recently, bridge continuous PDE and discrete max-flow min-cut graph methods. Presently we are surveying the spectral graph methods. Note that all segmentation methods are intrinsically probabilistic: we have spent a lot of our effort on further developing the NP Windows method of estimating probability density functions from a small number of samples of a signal/image.

Investigators:

Medical Vision Laboratory, Oxford University

Microsoft Research Cambridge

Posters:

Presentations

Abstracts:

  • Niranjan Joshi, Sarah Bond, and Michael Brady, "Segmentation of colorectal cancer MR images", 16th Scientific Meeting and Exhibition of International Society for Magnetic Resonance in Medicine - ISMRM, May 2008, Toronto, Canada.

Conference Papers:

Contact: