Current Projects

NIH/NCI/NIBIB R01 CA114492-01A2

High Performance Pattern Analysis for CT Colonography

Principal Investigator: Peter Santago

Section 1: Purpose

Background. Computed tomographic colonography (CTC) is a minimally-invasive screening method for detecting polyps. Mixed success has been achieved in clinical trials, in large part due to variations in reader experience, the large number of images, and the complex geometry of the colon. Yet computer polyp detection (CPD) and computer-aided polyp detection (CAPD) promise to improve the sensitivity and specificity of CTC, similar to mammography and lung nodule detection. Our long term goal is to further develop CAPD. Clinically effective CAPD with sufficient sensitivity and specificity requires CPD methods that utilize high-performance pattern analysis during development, training, and application.

Scope of Work.

• develop a state-of-the-art CPD system that utilizes high-performance pattern analysis
• extensively train the system using data from older scans, newer scans, and data from other institutions (e.g. Pickhardt et al. study).
• deploy the system and do performance study (new images from GCRC?)
• make the system available for different high-performance computing environments (clusters and grids)
• prepare preliminary results from other diseases for obtaining additional funding.

Strategy. Divide and conquer.

Section 2: Objectives

Aim 1. Design and implement a high-performance pattern analysis system for CPD. Computer identification of polyps is difficult at best and requires solutions to varied image analysis and optimization problems. Considering the complexity of the anatomy and the possibility of representing the colon in the computer in at least two distinct ways, vertex and voxel, a simple pattern recognition method is unlikely to meet the performance criteria. The ability of a single classifier type to perform optimally in this complex recognition task is summarized by the No Free Lunch Theorem (Wolpert and Macready, 1997): if one classifier performs better than another it is because it is better trained in a single aspect of the problem rather than being the best solution. Classifier ensembles offer a potential solution to this problem. For Specific Aim 1 we will

A. extend our current CPD system by incorporating more sophisticated
    pattern classification algorithms including and explicitly using both vertex
    and voxel representations;

B. incorporate these new algorithms into classifier ensembles;

C. use the classifier ensembles to validate automated feature identification
    and reduction;

D. incorporate a confidence metric into the CPD decision process; and

E. map the CPD system to a computer cluster.

Aim 2. Optimally combine prone and supine CTC scans. Current CPD and CAPD methods that acquire prone and supine do not fully utilize joint information. Colon segment registration within a patient study for polyp detection and across patient studies for polyp surveillance is a challenging problem as the colon undergoes significant deformation during repositioning. In addition, drastic local intensity changes caused by residual fluid movement complicate registration. For Specific Aim 2 we will

A. develop a non-rigid, hierarchical-based method to register the prone to
    supine colon lumen; and

B. jointly process the prone and supine scans and incorporate this method
    into the CPD system.

Design Objectives

Our goal is to have a flexible system that can be applied to other computer-aided diagnosis, implemented to take advantage of the different forms of high-performance computing.