Loading...
Publication
Aplicação de máquinas de vector suporte para classificação de ratos transgénicos através de imagem da retina
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 459.71 KB | Adobe PDF |
Authors
Abstract(s)
O objetivo deste trabalho consistiu na criação de modelos de aprendizagem
supervisionada baseados nas técnicas de Support Vector Machine (SVM) e Support Vector
Machine com informação privilegiada (SVM+) capazes de distinguir entre ratos saudáveis
(C) e transgénicos (D) por meio de análise de textura da imagem de tomografia de coerência
óptica (OCT) de retinas do olho direito.
A amostra é composta por 74 ratos, sendo 40 saudáveis e 34 transgénicos.
A tomografia de coerência óptica foi utilizada para obtenção da imagem da retina
dos ratos que, por sua vez, foi dividida em 4 quadrantes. A partir destes, obteve-se uma
imagem de fundo 2D e foram aplicados 20 indicadores de análise de textura de imagem de
fundo, usados como features para o modelo SVM.
As features com maior capacidade de separação entre grupos e que possuem
coeficiente de correlação inferior a 0,7 entre elas foram Inertia (primeiro, segundo e quarto
quadrantes), INN (Inverse difference normalized; terceiro quadrante), IMC2 (Information
measure of correlation; terceiro quadrante) e ClusterShade (terceiro quadrante).
Considerando as 6 features mais relevantes foram criados os modelos SVM e
SVM+ cujos parâmetros foram afinados de maneira a obter os modelos com a melhor
precisão na classificação dos ratos nas categorias saudável e transgénico. A técnica de
validação cruzada em 5 grupos foi utilizada para validar os resultados dos modelos criados.
Tanto para o conjunto de teste como para o conjunto de dados total o modelo SVM
obteve 100% precisão, enquanto que a precisão obtida pelo modelo SVM+ foi de 93,33%
(erro de apenas 1 caso em 15 – conjunto de teste) na classificação dos dados do conjunto de
teste e 98,65% (erro de apenas 1 caso em 74 – conjunto de dados total) no conjunto de dados
total.
The aim of this work was to create supervised learning models based on the Support Vector Machine (SVM) and Support Vector Machine with privileged information (SVM +) capable of distinguishing between healthy (C) and transgenic (D) mice through texture analysis of the optical coherence tomography (OCT) image of the retinas of the right eye. The sample consists of 74 mice, 40 healthy and 34 transgenic. Optical coherence tomography was used to obtain the image of the mice's retina, which in turn was divided into 4 quadrants. From these, a 2D background image was obtained and 20 background image texture analysis indicators were applied, used as features for the SVM model. The features with greater separation capacity between groups and which have a less than 0.7 correlation coefficient between each other were Inertia (first, second and fourth quadrants), INN (Inverse difference normalized; third quadrant), IMC2 (Information measure of correlation; third quadrant) and ClusterShade (third quadrant). Regarding the 6 most relevant features, the SVM and SVM + models were created, whose parameters were adjusted in order to obtain the models with the best precision in the classification of mice in the healthy and transgenic categories. The 5 fold crossvalidation technique was used to validate the results of the models created. For both, the test set and the total data set, the SVM model obtained 100% accuracy, while the precision obtained by the SVM + model was 93.33% (error of only 1 case in 15 - test set) in the classification of the test set data and 98.65% (error of only 1 case in 74 - total data set) in the total data set.
The aim of this work was to create supervised learning models based on the Support Vector Machine (SVM) and Support Vector Machine with privileged information (SVM +) capable of distinguishing between healthy (C) and transgenic (D) mice through texture analysis of the optical coherence tomography (OCT) image of the retinas of the right eye. The sample consists of 74 mice, 40 healthy and 34 transgenic. Optical coherence tomography was used to obtain the image of the mice's retina, which in turn was divided into 4 quadrants. From these, a 2D background image was obtained and 20 background image texture analysis indicators were applied, used as features for the SVM model. The features with greater separation capacity between groups and which have a less than 0.7 correlation coefficient between each other were Inertia (first, second and fourth quadrants), INN (Inverse difference normalized; third quadrant), IMC2 (Information measure of correlation; third quadrant) and ClusterShade (third quadrant). Regarding the 6 most relevant features, the SVM and SVM + models were created, whose parameters were adjusted in order to obtain the models with the best precision in the classification of mice in the healthy and transgenic categories. The 5 fold crossvalidation technique was used to validate the results of the models created. For both, the test set and the total data set, the SVM model obtained 100% accuracy, while the precision obtained by the SVM + model was 93.33% (error of only 1 case in 15 - test set) in the classification of the test set data and 98.65% (error of only 1 case in 74 - total data set) in the total data set.
Description
Keywords
Modelos de aprendizagem supervisionada SVM (Support Vector Machines) SVM+ Tomografia Imagem Retina Validação cruzada Features Supervised learning models SVM SVM+ Cross-validation Features