Comparação entre redes neurais convolucionais e técnicas de pré-processamento para classificar células HEp-2 em imagens de imunofluorescência
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
As doenças autoimunes são a terceira causa de mortalidade no mundo. A identificação de autoanticorpos nucleares por meio do teste de imunofluorescência indireta (IFI) em células epiteliais humanas (HEp-2 ) é um método convencional para auxiliar no diagnóstico e monitoramento de tais doenças. Entretanto, análises manuais de cé- lulas HEp-2 por IFI foram feitas ao longo dos anos, mas são subjetivas e demoradas. Portanto, a necessidade de métodos automatizados e padronizados é reconhecida e investigada há muito tempo e o desenvolvimento de sistemas de diagnóstico médico auxiliado por computador baseados em processamento de imagens e aprendizado de máquina se tornaram fundamentais. O principal objetivo deste trabalho consiste em avaliar cinco modelos de Redes Neurais Convolucionais (CNNs): LeNet-5, AlexNet, Inception-V3, VGG-16 e ResNet-50, para a tarefa de classificação automática de células HEp-2 e o impacto de seis estratégias de pré-processamento baseadas em alargamento de contraste, centralização de dados e aumento do conjunto de dados. Adicionalmente, a otimização de hiperparâmetros utilizando Tree of Parzen Esti- mators (TPE) é estudada junto a duas possíveis estratégias para explorar o poder das CNNs em diferentes cenários: treinamento completo e fine-tuning. Todos os ex- perimentos foram realizados utilizando o conjunto de dados HEp-2 do ICPR 2014 composto por 13.596 imagens categorizadas em 6 classes. O desempenho da classi- ficação foi avaliado por meio da validação cruzada estratificada com 5-fold sobre o conjunto de treino e a comparação entre os modelos CNN considerou o conjunto de teste. O melhor resultado, em termos de acurácia, foi alcançado treinando o modelo Inception-V3 a partir do zero, sem pré-processamento e com aumento de dados, com valores situados em 98,28%, que supera os resultados apresentados por outros tra- balhos na literatura. Os resultados apontaram que a maioria das CNNs tem melhor desempenho sobre imagens não pré-processadas quando treinadas a partir do zero e com aumento de dados. Embora a técnica de fine-tuning pareça promissora, ela não obteve resultados melhores do que as CNNs treinadas a partir do zero. Talvez, isso seria considerado em cenários em que o tempo de treinamento é um problema, pois a técnica exige menor tempo computacional. Os métodos desenvolvidos neste trabalho podem auxiliar os especialistas em saúde a escolherem corretamente um método de classificação para identificar e acompanhar condições autoimunes e as doenças que as causam.
Autoimmune diseases are the third cause of mortality in the world. The identification of antinuclear antibody via indirect immunofluorescence (IIF) test in Human Epithelial-2 (HEp-2) cells is a conventional method to support the diagnosis and monitoring of such diseases. However, manual analyses of HEp-2 cells by IIF have been done along years, but it is subjective and time-consuming. The need for au- tomated and standardized methods is known and persecuted for a long time, and the development of computer-aided diagnosis systems based on image processing and machine learning techniques are fundamental. The main objective of this rese- arch resides in assess five Convolutional Neural Networks (CNNs) models: LeNet-5, AlexNet, Inception-V3, VGG-16 and ResNet-50, for this task of automatic classifi- cation of HEp-2 cells and the impact of six different pre-processing strategies based on contrast improvements, data centralization, and data augmentation. Additio- nally, the hyperparameters optimization using Tree of Parzen Estimators (TPE) is applied, and two possible strategies for exploiting the power of existing CNNs in dif- ferent scenarios are analyzed : from scratch and fine-tuning. All experiments were performed using the HEp-2 dataset from ICPR 2014 composed by 13,596 images classified in six different classes. The classification performance was evaluated using stratified 5-fold cross-validation over the training set, and the comparison among the CNN models considered the test set. The best result, in terms of accuracy, was achieved by training the Inception-V3 model from scratch without preprocessing and with data augmentation, with values lying on 98.28%, which outperforms the results presented in other works in literature. The results pointed out that most of CNNs perform better over non-preprocessed images when trained from scratch and data augmentation. Although fine-tuning technique sounds promising, it did not achieve better results than the CNNs trained from scratch. Maybe, it would be con- sidered in scenarios where training time is an issue, as its computational needs are lower. The methods developed in this research can help health agents to correctly choose a classification method to identify and manage autoimmune conditions and the diseases that cause them.
Autoimmune diseases are the third cause of mortality in the world. The identification of antinuclear antibody via indirect immunofluorescence (IIF) test in Human Epithelial-2 (HEp-2) cells is a conventional method to support the diagnosis and monitoring of such diseases. However, manual analyses of HEp-2 cells by IIF have been done along years, but it is subjective and time-consuming. The need for au- tomated and standardized methods is known and persecuted for a long time, and the development of computer-aided diagnosis systems based on image processing and machine learning techniques are fundamental. The main objective of this rese- arch resides in assess five Convolutional Neural Networks (CNNs) models: LeNet-5, AlexNet, Inception-V3, VGG-16 and ResNet-50, for this task of automatic classifi- cation of HEp-2 cells and the impact of six different pre-processing strategies based on contrast improvements, data centralization, and data augmentation. Additio- nally, the hyperparameters optimization using Tree of Parzen Estimators (TPE) is applied, and two possible strategies for exploiting the power of existing CNNs in dif- ferent scenarios are analyzed : from scratch and fine-tuning. All experiments were performed using the HEp-2 dataset from ICPR 2014 composed by 13,596 images classified in six different classes. The classification performance was evaluated using stratified 5-fold cross-validation over the training set, and the comparison among the CNN models considered the test set. The best result, in terms of accuracy, was achieved by training the Inception-V3 model from scratch without preprocessing and with data augmentation, with values lying on 98.28%, which outperforms the results presented in other works in literature. The results pointed out that most of CNNs perform better over non-preprocessed images when trained from scratch and data augmentation. Although fine-tuning technique sounds promising, it did not achieve better results than the CNNs trained from scratch. Maybe, it would be con- sidered in scenarios where training time is an issue, as its computational needs are lower. The methods developed in this research can help health agents to correctly choose a classification method to identify and manage autoimmune conditions and the diseases that cause them.
Description
Citation
RODRIGUES, Larissa Ferreira. Comparação entre Redes Neurais Convolucionais e técnicas de pré-processamento para classificar células HEp-2 em imagens de imunofluorescência. 2018. 65 f. Dissertação (Mestrado em Ciência da Computação) - Universidade Federal de Viçosa, Viçosa. 2018.
