Robust 3D Segmentation of Pulmonary Nodules in Multislice CT Images

Kazunori Okada, Dorin Comaniciu
Real-Time Vision and Modeling Department, Siemens Corporate Research, Inc.
kazokada@sfsu.edu
dorin.comaniciu@siemens.com
Arun Krishnan
CAD Program, Siemens Medical Solutions USA, Inc.
arun.krishnan@siemens.com

We propose a robust and accurate algorithm for segmenting the 3D pulmonary nodules in multislice CT scans. The solution unifies i) the parametric anisotropic Gaussian model fitting of the volumetric data evaluated in Gaussian scale-space and ii) non-parametric 3D segmentation based on normalized gradient (mean shift) ascent defining the basin of attraction of the target tumor in the 4D spatial-intensity joint space. This unification, by treating the parametric estimation results from the first step as a normal prior, realizes an efficient 3D tumor segmentation according to both spatial and intensity proximities simultaneously. Experimental results show that the system reliably segments a variety of nodules including part- or non-solid nodules which poses difficulty for the existing solutions. The system also efficiently processes a 32x32x32-voxel volume-of-interest by six seconds on average.


Our method's segmentation results are shown below. We use a data set of 14 patients with 77 nodules whose size ranges from 3mm to 25mm in diameter. Examples are shown for i) part- or non-solid nodules, ii) pleural attached nodules, iii) vascularized nodules, and iv) small nodules (~3mm).

For comparison, the results by a) the parametric (anisotropic Gaussian/ellipsoidal) model fitting (2nd columns) and b) the non-parametric 4D joint space segmentation (3rd columns) are shown for each nodule cases. In the 2nd columns, "+" and "x" indicate the marker location and tumor center estimate, respectively. In the 4th columns, the white pixels outline the outside boundary of the segmented tumors. The 3rd columns show the 3D Gaussian density estimated from the first step of our system. The 5th columns show the 3D projection of the 4D joint space density estimates from the second step.




(i) Part- or Non-Solid Nodules


3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy






3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy








(ii) Pleural Attached Nodules


3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy






3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy








(iii) Vascularized Nodules


3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy






3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy








(iv) Small Nodules


3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy






3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy








(v) Examples


3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy






3D Input
3D Ellipoidal Fit
3D Normal Density Estimate
3D Segmentation Result
4D Joint Density Estimate
yz





xz





xy