Associação entre atributos manuais e aprendizado profundo baseada em geometria fractal para classificação de imagens histológicas

Abstract

Classification of histology images is a task that has been widely explored on recent computer vision researches. The most studied approach for this task has been the application of deep learning through CNN models. However, the use of CNN in the context of histological images classification has yet some limitations such as the need of large datasets and the difficulty to implement a generalized model able to classify different types of histology tissue. In this project, an ensemble model based on handcrafted fractal features and deep learning that consists on combining the classification of two CNN and the classification of local and global handcrafted features by applying the sum rule is proposed. Feature extraction is applied to obtain 300 fractal features from different histological datasets. These features are reshaped into a matrix in order to compose an RGB feature image. Four different reshaping procedures are evaluated, wherein each generates a representation model of fractal features which is given as input to a CNN. Another CNN receives as input the correspondent original image. After combining the results of both CNN with the classification of the handcrafted features using classical machine learing approaches, accuracies that range from 88.45\% up to 99.77\% on five different datasets were obtained. Moreover, the model was able to classify images from datasets with imbalanced classes, without the need of images having the same resolution, and using 10 epochs for training. It was also verified that the obtained results are complementary to the most relevant studies recently published in the context of histology image classification.