Transfer learning improves supervised image segmentation across imaging proto...
CREOSA_FORSON_Kofi
1. Classification of LIBS Spectra using Classificatory Decomposition: Leave-One-Out Approach
Kofi Forson, Kyrhee Powell, and Dr. Tomasz G. Smolinski
Department of Computer and Information Sciences, Delaware State University, Dover, DE
19901
In years prior, different methods have been applied to classify proteins from laser-induced
breakdown spectroscopy (LIBS) data. Some of these proteins include Bovine Serum Albumin,
Opsteopontin, Leptin, and Insulin-like Growth Factor II. The classification of these particular
proteins can lead to the detection of diseases, such as ovarian cancer. We hypothesize that
classificatory decomposition (CD) is an effective method to classify LIBS data into the four
protein types. CD uses multi-objective evolutionary algorithms and rough sets to classify data.
This method not only decomposes the spectra into a small set of additive components by using
multi-objective optimization, but also tests the classificatory aptitude of the decomposition.
The goals of classificatory decomposition are to obtain a low reconstruction error rate and high
classification accuracy, while utilizing few components. Classificatory decomposition uses
pareto-optimality, reducts, and elitist non-dominated vector evaluated genetic algorithms (end-
VEGA, NSGA II) to classify LIBS data into the protein types and reach these goals. We will be
using two different approaches called Cross Validation and Leave-One-Out to evaluate the
accuracy of the data. Cross Validation involves splitting the data into ten different sections,
taking one section out, training the learning scheme on the remaining nine sections, and then
testing the scheme on the section that was taken out. I will focus on the Leave-One-Out
approach, which involves taking one instance out, training the learning scheme on the
remaining instances, then testing the scheme on the instance that was taken out. The
components used to classify the spectra can also be reused to classify future LIBS data. Further
research could entail applying classificatory decomposition to spectra that are not necessarily
of the same size.
Funder Acknowledgement: This study was supported by the NSF CREST Center for Research
and Education in Optical Sciences and Applications (CREOSA), HRD 1242067.
Presentation Mode: Poster