Subject: Visiting talk, March 15th (Thursday)

Title:ÊÊÊ A Geometry-Based Method for Colon Polyp Detection
Ê
Presenter: ÊPadma Sundaram, PhD, Stanford University, Electrical Engineering Department
Ê
When: March 15th Ê(Thursday), 11:00am-noon
Ê
Where: ÊNuclear Medicine Conference Room (Building 10)
Ê
Abstract:
Ê
Colon cancer is the second leading cause of cancer deaths in the United States. Polyps are small lumps on the colon surface that are considered to be precursors to colon cancer. Early detection and removal of polyps may help reduce the risk of colon cancer. Radiologists typically detect polyps by inspecting computed tomographic (CT) images of the colon.
However, the method is inherently time consuming as thousands of CT images must be individually interpreted, also leading to the possibility of fatigue induced error.
Ê
Computer-aided detection (CAD) algorithms may help reduce interpretation time and improve sensitivity. The input to a CAD algorithm is the 3-d image obtained from the CT scan of the colon. This is a voxelized image containing density values. The aim of a CAD algorithm is to identify locations in the colon that are most likely to contain polyps. Since an idealized polyp is a curved hemispherical protrusion of the colon surface, algorithms may attempt to detect polyps through their shape.
Also, since the primary characteristic of shape is curvature, any method that detects polyps through their shapes must estimate curvature, directly or indirectly. Unfortunately, curvature is a smooth notion, while our data are discrete and noisy. Moreover, as a second order differential quantity, curvature amplifies noise, resulting in unstable and noisier estimates.
Ê
Nevertheless, curvature forms the basis of almost all existing algorithms in the literature. While different in detail, these methods process the image volume as a grid of regularly arranged samples containing density values. They compute curvature by convolution of the voxel data with appropriate kernels that are aligned to the voxel grid.
This results in the inclusion of Òirrelevant but nearbyÓ data in the computations. This is because treating the data as a voxel grid discards any topology information present in the images. This is important when the object imaged is a complex structure like the colon and is closely surrounded by unrelated structures. Curvature being a second order differential quantity, any noise (structural or image) is amplified, resulting in unstable, noisy curvature estimates.
Ê
In contrast, in this work, the data is processed in the form of a discrete mesh. We compute curvature on the mesh using techniques explicitly designed for piecewise linear surfaces. The curvature computation process does not assume locally smooth manifolds. The discrete mesh is assumed to be the only given thing. All subsequent computations take place ÒonÓ the mesh. Since the computations are restricted to the mesh surface, the resulting curvature estimates are stable and unaffected by the presence of nearby, irrelevant structures.
I will discuss the advantages of this approach and also present an evaluation of the algorithm on a patient database. I will also address the effect of CT acquisition parameters on algorithm performance.
ÊÊ
for additional information contact Jack Yao, PhD 301-402-3225
Ê