Abstract Archives of the RSNA, 2003
K19-1030
Computerized Scheme for Distinction between Benign and Malignant Nodules in Thoracic Low-dose CT by Use of Massive Training Artificial Neural Network
Scientific Papers
Presented on December 3, 2003
Presented as part of K19: Physics (Image Processing: CAD V--Lung)
Kenji Suzuki PhD, PRESENTER: Nothing to Disclose
Abstract:
HTML
Purpose: Low-dose helical CT (LDCT) is being used as a useful tool for lung
cancer screening. It is very difficult, however, for radiologists to
distinguish between benign and malignant nodules in LDCT. Our purpose in this
study is to develop computer-aided diagnostic (CAD) scheme for distinction
between benign and malignant nodules in LDCT scans by use of a novel pattern-classification
technique based on a massive training artificial neural network (MTANN).
Methods and Materials: We developed a multiple MTANN that consists of three
expert MTANNs for distinction between malignant nodules and three groups of
benign nodules. Each of the MTANNs was trained by using input CT images and the
desired output images containing the distribution for the "likelihood of
being a malignant nodule." The MTANNs were massively trained by using a
large number of overlapping sub-regions obtained from the input image. Each
MTANN was trained by using ten typical malignant nodules and ten benign nodules
representing a specific type of benign nodule in each of three groups, i.e.,
each MTANN was applied for distinguishing malignant nodules from (1) small
benign nodules with relatively high contrast, (2) benign nodules in the
peripheral region, and (3) other types of benign nodules. The three MTANNs were
combined by using an integration ANN such that these three groups of benign
nodules can be distinguished from malignant nodules. Our database consisted of
72 primary lung cancers in 69 patients and 412 benign nodules in 337 patients,
which were obtained from a lung cancer screening program on 7,847 screenees
with LDCT for three years in Nagano, Japan. All cancers were confirmed
surgically, and all benign nodules were diagnosed based on follow-up CT
examinations. Another database consisted of 15 "missed" lung cancers
that were not reported as lung cancers due to interpretation errors during the
initial clinical reading. The performance of the multiple MTANN was evaluated
by using receiver operating characteristic (ROC) analysis.
Results: Our scheme achieved an Az (area under the ROC curve) value of 0.87 in
the round robin test, whereas an average Az value of 0.70 was obtained by 16
radiologists in our observer study. Our scheme identified 97.2% (70/72) of
malignant nodules, and 86.7% (13/15) of interpretation error cases correctly as
malignant, whereas 49.3% (203/412) of benign nodules were identified correctly
as benign.
Conclusion: Our computerized scheme for distinction between benign and
malignant nodules would be useful in assisting radiologists in diagnosis of
lung nodules. (K.D. is a shareholder in R2 Technology, Inc., Sunnyvale, CA, and
Deus Technologies, Inc., Rockville, MD.)
Questions about this event email: suzuki@uchicago.edu
Suzuki PhD, K,
Computerized Scheme for Distinction between Benign and Malignant Nodules in Thoracic Low-dose CT by Use of Massive Training Artificial Neural Network. Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL.
http://archive.rsna.org/2003/3100417.html