Curvature Analysis Based Framework for Virtual Colonoscopy
Dongqing Chen · 2009
This dissertation is concerned with the development of a novel general framework for polyp detection and visualization using a unique color, image pre-processing techniques, and clinical validation of fly-over navigation in VC. Since the colonic polyps grow from mucosa (the inner layer of colon) into the lumen, they are modeled as protrusion shapes. A more general criterion, which is based on Gaussian curvature K = κ1κ2 > 0 rather than the minimum principal curvature κ2 > 0, provides more powerful detection abilities on convex and concave protrusion shapes at the same time. The important feature of the proposed model is that it can detect protrusions with both convex and concave shapes. Protrusion shapes are defined as the extension beyond the usual limits or above a plane surface. Based on Gaussian and mean curvature flows, the approach works by locally deforming the convex or concave surface until the second principal curvature goes to zero. The diffusion directions are changed to prevent convex surfaces from converting into concave shapes, and vice versa. The deformation field quantitatively measures the amount of protrudeness. The proposed method has been evaluated by using synthetic phantoms and real colon datasets. First, a 3D colon model is generated by searching the iso-surface using the Marching Cubes algorithm. The iso-surface is triangulated into a mesh surface, where the polyp candidates are found by the proposed surface evolution algorithm and three additional geometry features. Then, all the detected polyps as well as their neighborhoods are placed at the same locations of a second polygonal dataset with the same topology and geometry properties as the triangle mesh surface. Finally, different colors are assigned to the two separated datasets to highlight the polyp candidates to enhance the visualization effect during VC navigation. The image pre-processing technique for clinical colon datasets mainly includes colon tissue segmentation, 3D colon iso-surface generation and rendering, and 3D colon lumen centerline extraction. First a simple thresholding method is used to remove most of the opacified liquid. Most opacified liquid is roughly removed, and some tissues such as bones on the same slice are removed as well, since they have the similar image intensities. However, it does not affect the colon segmentation at all, since we manually select the initial seeds totally inside the colon region, which guarantees the segmentation results. Then, closed contours with manual seed initialization are propagated toward the desired region boundaries through the iterative evolution of an adaptive implicit function. During the iterations, every candidate pixel is automatically classified to either colon part or background based on its probability density function (PDFs), by using Bayesian decision criterion. After each iteration, the average values and variances for both foreground and background are estimated by using the Maximum Likelihood (ML) method respectively. The iteration procedure will stop when level sets curves converge to the lumen-air boundaries. Finally, the accuracy of the proposed method is evaluated by computing the overlaps between the manually segmented colon and the results segmented by algorithm. The common visualization technique in virtual colonoscopy that tries to simulate the real one is the virtual fly-through navigation. According to [1] a clinical study has shown that performing fly-through navigation in one direction only (antegrade or retrograde), covers 61%-91% with an average of 80% of the colon surface depending on the field of view of the camera. The Fly-Over visualization technique is firstly proposed to address this problem by our group in 2006 [2], which is designed to overcome the limitations of existing techniques. The main idea is to split the entire colon anatomy into exactly two halves, and assign a virtual camera to each half to perform fly-over navigation. This visualization technique has been evaluated, and the whole evaluation procedure has been conducted on the 10 clinical datasets to test fly-over versus fly-through navigation. The comparison results show that the average surface visibility of fly-over visualization is 97.49±1.02%, which is higher those of one-way fly-through visualization at 77.64±4.86% and two-direction fly-through visualization at 93.40±1.21%. (Abstract shortened by UMI.)