Abstract:
HTML
Purpose: To develop a lung contrast normalization method to minimize the
effect of variation in image properties of digital chest images obtained from
different systems on the performance of computer-aided detection (CAD) for
early-stage lung cancers.
Methods and Materials: RapidScreen RS-2000 is the only FDA approved commercial
CAD system for lung cancer detection based on chest radiography. This
film-based system is trained on a database consisting of over 10,000 cancer and
cancer-free cases. To apply the existing CAD to digital images obtained from
computed radiography (CR) and direct digital radiography (DR) systems of
various vendors, we developed a normalization method to overcome the influence
of large variation in image properties including image contrast, gray scale and
pixel resolution. For this study, CR and DR images are obtained from systems
made by 6 and 2 different manufactures, respectively. For all of these images,
the pixel size varies from 0.139 to 0.2 mm, and gray scale from 10 to 14 bits.
The relative lung contrast also has a variation of more than 500 pixel values
among them. Our method consists of two steps, namely pixel resolution
normalization and contrast gray scale normalization. The pixel size of an input
image is normalized to 0.7 mm by reducing the image matrix size with box
filtering. The differences between the maximum and minimum pixel value in a
region at the image center is defined as the lung contrast. The pixel values of
input images are then normalized to 10-bit based on the inverse relationship to
the defined contrast to achieve a uniform lung contrast among images. CAD
nodule detection is then applied to normalized images.
Results: In clinical trial studies, 80 chest films and 79 digital images with
cancers in lungs are used to compare the detection performance. The sensitivity
and average number of false positive per image (ANFP) are 66.3% and 5.0/image
for films, respectively. If the normalization is not used, the sensitivity for
digital images is reduced by more than 25% with a similar ANFP. However, the
sensitivity is improved to 63.3% with the normalization. Between the
performance of films and normalized images, the estimated differences and its
95% confidence interval for the sensitivity and ANFP are 0.03 and (-0.13 and
0.19), and 0.0005 and (-0.5, 0.5), respectively.
Conclusion: The results indicate the RS-2000 can be successfully applied to
various digital images with normalization for early-stage lung cancers. This
allows the RS-2000 to be integrated in PACS systems and fitted into
radiologist's workflows.
(X.X., F.L. are employees of Deus Technologies. M.Y. is owner of Deus
Technologies. M.F., B.L. are shareholders in Deus Technologies.)
Xu PhD, X,
Effectiveness of Lung Contrast Normalization to Automatic Detection of Lung Nodules on Digital Chest Images Obtained from Different Types of Acquisition Systems. Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL.
http://archive.rsna.org/2003/3101307.html