Current Research

Pulmonary Nodule Analysis on CT scans

Artit Jirapatnakul

From the I-ELCAP website:

Lung cancer is the leading cause of death from cancer in both men and women in the United States, killing more people than cancers of the breast, colon, cervix and prostate combined. In the year 2005, it was estimated that there would be over 163,510 deaths in the United States alone; worldwide, more than a million die from it every year. Lung cancer has one of the worst prognoses of all cancers, with an overall cure rate of 5%...

Lung cancer usually does not cause any symptoms until it has reached an advanced stage, when treatment is least likely to work. Early Stage I lung cancer discovered through early detection screening, however, has a cure rate of 70%, and for some subgroups it is even higher.

Lung caner often first manifests as opaque lesions in the lung, sometimes referred to as "pulmonary nodules". Recent advances in helical CT scanner technology have improved the resolution of the CT images, along with the speed at which the images are acquired. The increase in resolution in both the in-plane and out-of-plane resolutions introduces not only larger images, but a larger number of slices. This poses an additional burden on the radiologist because she/he must not only review larger and more detailed images; but she/he must also review a larger number of images than before. With such an information explosion, it becomes natural to use a computer as a tool to assist the radiologist in the detection, identification, analysis, cataloging, visualization and diagnosis of pulmonary nodules in CT scans.

Automated analysis of pulmonary nodules can be divided into two main areas: measurement and characterization. The ultimate goal is to incorporate algorithms in both areas into radiological workstations for use by radiologists in clinical practice

Measurement

One of the most reliable indicators of the malignancy of pulmonary nodules is their growth rate, with higher growth rate typically more indicative of cancers. Accurate growth rate measurement requires accurate and robust nodule segmentation algorithms. A recent publication describes the semi-automated method developed by Reeves et al on nodule segmentation [1]. Briefly, when provided with the location of a nodule in a CT scan, the algorithm extracts a region around the nodule, segments the nodule from attached structures, and produces colored two-dimensional and 3D images of the segmented nodule. Volume is computed from the segmented image.

Progress is being made to improve the reliability of the segmentation algorithm in the presence of other attached structures, such as blood vessels, airways, and the thoracic wall. One difficulty in the evaluation of any segmentation or measurement algorithm for pulmonary nodules is that it is impossible to determine the "true" nodule size. Recent work suggests that there is large deviation even amongst radiologists as to the size of a nodule [2]. One possible substitute is the use of mathematical models to estimate the nodule volume based on previous scans of the same nodule [3]. An exponential model is commonly used to describe the growth of malignant nodules; however, fitting such a model requires at least three comparable scans of the same nodule, which are often difficult to obtain.

Characterization

In contrast to growth analysis, automated methods for nodule characterization use features available from one CT scan to determine the malignancy of the nodule. Some features that might be used include size, shape, or density. Characterization is a difficult task due to the lack of knowledge of what attributes are correlated with malignancy.

Preliminary results on characterizing nodules using 3D shape, curvature, and density features were reported by Jirapatnakul et al [4]. There have been many other studies in this area, but for many of the studies for which size-distribution information was available, much of the reported performance could be derived from the size of the nodules alone [5, 6]. Currently, work is focused on the development of automated nodule characterization systems that are able to significantly improve upon the performance of size alone and the creation of a dataset where the size-distribution of malignant and benign nodules is less skewed.

References

1. A. Reeves, A. Chan, D. Yankelevitz, C. Henschke, B. Kressler, and W. Kostis.
   On measuring the change in size of pulmonary nodules
   IEEE Transactions on Medical Imaging, , 25: 435-450, 2006.
2. A. P. Reeves, A. M. Biancardi, T. V. Apanasovich, C. R. Meyer, H. MacMahon, E. J.R. van Beek, E. A. Kazerooni, D. Yankelevitz, M. F. McNitt-Gray, G. McLennan, S. G. Armato III, C. I. Henschke, D. R. Aberle, B. Y. Croft, and L. P. Clarke. "The lung image database consortium (LIDC): A comparison of different size metrics for pulmonary nodule measurements." Academic Radiology, 14(12):1475-1485, Dec 2007.
3. A.C. Jirapatnakul, A. P. Reeves, T. V. Apanasovich, M. D. Cham, D. F. Yankelevitz, and C. I. Henschke. Characterization of solid pulmonary nodules using three-dimensional features, SPIE International Symposium on Medical Imaging, 2007.
4. A.C. Jirapatnakul, A. P. Reeves, T. V. Apanasovich, M. D. Cham, D. F. Yankelevitz, and C. I. Henschke. Prediction of tumor volumes using an exponential model, SPIE International Symposium on Medical Imaging, 2007.
5. A. Jirapatnakul; A. P. Reeves, T. V. Apanasovich, A. Biancardi, D. F. Yankelevitz, and C. I. Henschke. Pulmonary Nodule Classification: Size Distribution Issues, 2007. IEEE International Symposium on Biomedical Imaging: From Macro to Nano, 2007.
6. A. C. Jirapatnakul, A. P. Reeves, T. V. Apanasovich, A. M. Biancardi, D. F. Yankelevitz, and C. I. Henschke. "Characterization of pulmonary nodules: Effects of size and feature type on reported performance," In M. L. Giger and N. Karssemeiger, eds., SPIE International Symposium on Medical Imaging, vol. 6915, p. 69151E, Feb 2008.

List of Current Research Projects