Abstract:
HTML
Purpose: To investigate the use of three- and four-dimensional
non-parametric classifiers of mammographic lesions for use in computer-aided
diagnosis. Because the non-parametric density estimation is asymptotically
unbiased, this classifier is asymptotically optimal. We compared the
performance of this classifier and of Linear Discriminant Analysis (LDA) for a
finite sample size.
Methods and Materials: We investigated a non-parametric classifier, which takes
as inputs computer-extracted features of mammographic lesions and outputs a
numerical value related to the probability of malignancy. A non-parametric
classifier consists of two steps. First, non-parametric density estimations
(NPDE) of the density distributions of malignant and benign lesions are determined.
For each multidimensional space corresponding to a given set of
computer-extracted features, the estimated density distribution is the sum of
blurring kernels (dome shaped function with unit integral) centered on all the
observations. The density distribution is computed for both benign and
malignant lesions from a training set. In the second step, lesions from a
testing set are classified by computing an estimate of the likelihood ratio as
the ratio of the two density estimates at the point corresponding to each
unknown lesions' characteristics. The dimensionality of the classifier
corresponds to the dimensionality of the kernel. We used three and four
dimensional Gaussian and parabolic kernels. The training database consisted of
296 lesion images from 142 patients (148 benign and 148 malignant). The
independent testing database consisted of 204 lesion images from 102 patients.
Each lesion was automatically extracted from the parenchymal background and
five computer-extracted features were obtained for each lesion. ROC analysis
and Az value were used to assess the performance of the technique for
distinguishing between malignant and benign lesions for each feature
combination. Linear discriminant analysis (LDA) was also performed on the two
datasets.
Results: Az values were similar for the parabolic kernel, the Gaussian kernel,
and the LDA (within +/- 0.03, 0.002 on average). The four-feature combination
of radial gradient of margin, spiculation, margin sharpness, and average gray
value with the parabolic kernel produced the highest Az value of 0.78 on the
testing dataset.
Conclusion: For a finite sample size in a multi-dimensional setting, NPDE
performed as well or better than LDA. NPDE is a feasible alternative to LDA for
mammographic lesion classification.
(M.L.G. is a shareholder in and received a grant from R2 Technologies. L.L. is
a shareholder in R2 Technologies. )
Bonta MD, PhD, D,
Multi-dimensional Non-Parametric Classifier of Mammographic Lesions. Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL.
http://archive.rsna.org/2003/3103198.html
