Saved in:
| Main Author: | |
|---|---|
| Format: | Artículo científico |
| Language: | pt |
| Published: |
Universidade Nove de Julho
2006
|
| Subjects: | |
| Online Access: | https://www.redalyc.org/articulo.oa?id=81040211 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- Classification of mammographic features using RBF-SA Rafael do Espírito Santo Roseli de Deus Lopes Rangaraj M. Rangayyan Ingeniería Mammography Performance Optimization Neural networks Simulated annealing We present in this work a new type of classes discriminatorbased upon nonlinear and combinational optimization techniques:radial basis functions-simulated annealing (RBF-SA).The combinational optimization method is used here as a preestimationof some parameters of the network classifier. Wecompare the classifier performance with and without pre-estimation.For training the classifiers, adopting the leave-one-outprocedure, we have used case examples such as mammographicmasses (malignant and benign). The classifier is trained withshape factors and edge-sharpness measures extracted from 57regions of interest (ROI) (37 malignant and 20 benign), manuallydelineated, that describe mammographic masses and tumorfeatures in terms of polygonal models for shape factors (compactness[CC], Fourier description [FF], fractional concavity[FCC] and speculated index [SI]) and edge sharpness-acutance(A) . The classifier performance is compared in terms of thearea under the receive operating characteristic (ROC) curve (A). Higher values of A correspond to a better performance ofclassifier. Experiments with mammographic tumor and massesshow that the best result of 0.9776 is obtained with RBF-SAwhen RBF parameters such as centers and spread matrix arepre-estimated, which is significantly better than the resultsobtained with no pre-estimation or only pre-estimation of theRBF centers, which are, 0.7071 and 0.9552 respectively. 2006 artículo científico 1678-5428 https://www.redalyc.org/articulo.oa?id=81040211 pt http://www.redalyc.org/revista.oa?id=810 Exacta application/pdf Universidade Nove de Julho Exacta (Brasil) Num.2 Vol.4