Exploração de transformada Curvelet com redes neurais convolucionais para classificação de imagens histológicas

• Main Author: Gervaes, Guilherme
• Publication Date: 2021
• Format: Bachelor thesis
• Language: por
• Source: Repositório Institucional da UNESP
• Download full: http://hdl.handle.net/11449/216091
• topic: Feature extraction; Curvelet transform; Extração de características; Transformada curvelet; ResNet50

Feature extraction and the method utilized for data processing are fundamental parts of the digital image classification process. Therefore, in this work, the classification capacity of a convolutional neural network was observed when curvelet coefficients extracted from histological images were added to the original histological images, in contexts of colorectal cancer, non-Hodgkin lymphomas, breast cancer and liver tissue of mice of both genders and at different stages of life in order to observe the best combinations of histological images and curvelet coefficients in the classification process. In the feature extraction step, curvelet coefficients of histological images were obtained and sub-images composed of these coefficients were generated. The classification, performed by the pre-trained residual convolutional neural network ResNet50, used the images obtained in the feature extraction step combined with the classification using only histological images. The most promising results were observed when the sub-images generated from the coefficients were added to the original histological images, producing results up to 10.70% superior in accuracy and 25% superior in precision when compared to those observed in the classification using only histological images. The combination of the original histological images with the sub-images generated from the curvelet coefficients presents a relevant contribution for researchers interested in the development of computer-aided systems that help professionals in decision-making in the domains presented in the work.