Three-dimensional imaging is a well-established method in the diagnosis of diseases of the airway, the colon and the arterial System. In conjunction with fiberoptic methods and computed tomography (CT) three-dimensional imaging and reconstruction has proved to be useful in detection of tumors, obstructions, strictures and certain inflammatory lesions. Virtual endoscopy is a new method of displaying three-dimensional reconstruction of hollow anatomic structures that simulate conventional endoscopy. In contrast to conventional endoscopy, virtual endoscopy is completely noninvasive. The main approach uses the analysis of local differential geometry features, to automatically detect and assess masses of the airways, and the detection of colonic cancers and polyps, and a variety of local curvature measures have been introduced and applied to both patient and artificial data reconstructions. These functions derive mainly from the estimated values of the principal curvatures (mean and Gaussian curvature, local sphericity, wall thickness, polyp density, etc.) and automated detection schemes have been studied that invoke filter sets derived from the local geometry. We continue to work at improving sensitivity and specificity, now using a new, large patient data set, part of an on-going clinical trial at the Mayo clinic. Current detection strategy uses the combination of radiologist and CT scans, to conduct computer-aided detection. The data presented to the radiologist are pre-processed using filters derived from the curvature measures described above and then sent to either a neural net program, or to binary classification tree program, both trained specifically on these data. Either method was shown to significantly improve the sensitivity and specificity rates above those for detection using the radiologist or the computer-based analysis separately. Currently we have improved detection by means of ensemble methods (majority vote across a collection of base classifiers), and by using support vector machines as base classifiers. We also have significantly improved the estimation of error rates using the 632+ smoothed bootstrap approach (Efron and Tibshirani, JASA, 1997). We are now implementing a boosting scheme (interative re-weighting across many base classifiers) and an alternative smoothed bootstrap method for error estimation (which is known to reduce both bias and variance of the error estimates). [1] Multi-network Classification Scheme for Detection of Colonic Polyps in CT Colonography Data Sets, Anna K. Jerebko, James D. Malley, Marek Franaszek, Ronald M. Summers (in revision, Academic Radiology, 2002)
| Jerebko, Anna K; Summers, Ronald M; Malley, James D et al. (2003) Computer-assisted detection of colonic polyps with CT colonography using neural networks and binary classification trees. Med Phys 30:52-60 |
| Jerebko, Anna K; Malley, James D; Franaszek, Marek et al. (2003) Multiple neural network classification scheme for detection of colonic polyps in CT colonography data sets. Acad Radiol 10:154-60 |
| Summers, Ronald M; Jerebko, Anna K; Franaszek, Marek et al. (2002) Colonic polyps: complementary role of computer-aided detection in CT colonography. Radiology 225:391-9 |
| Summers, R M; Beaulieu, C F; Pusanik, L M et al. (2000) Automated polyp detector for CT colonography: feasibility study. Radiology 216:284-90 |
