Métodos computacionais baseados em superpixels para segmentação automática da próstata em imagens de ressonância magnética 3D

• Autor(a) principal: SILVA, Giovanni Lucca França da
• Data de Publicação: 2021
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFMA
• Texto Completo: https://tedebc.ufma.br/jspui/handle/tede/tede/3676
• topic: segmentação da próstata;; superpixels;; convolutional neural network;; índices filogenéticos;; algoritmo xgboost;; otimização por enxame de partículas;; prostate segmentation,; superpixels;; convolutional neural networks;; phylogenetic indexes;; xGBoost algorithm;; particle swarm optimization; Análise de Algoritmos e Complexidade de Computação

Prostate cancer is the second most common cancer in men in the world. In Brazil, there are an estimated 65,840 new cases of prostate cancer for each year of the 2020-2022 triennium. The automatic segmentation of the prostate is an important factor to assist in the diagnosis and treatment of cancer, such as orientation of the biopsy procedure and radiotherapy. However, automatic segmentation is challenging due to the great variation in the anatomy of the prostate due to pathological changes, tissue similar to Organs adjacent organs and different image acquisition protocols. Therefore, this work proposes three computational methods based on superpixels for automatic segmentation of the prostate in 3D magnetic resonance (MR) images. All the proposed methods consider the following steps: 1) description of the materials, 2) prostate detection, 3) image enhancement, 4) prostate segmentation, 5) refinement of the segmentation, and 6) evaluation of the results. The differences between the proposed methods are found in the segmentation of the prostate with the subset of superpixels classification. The first proposed method presents a classification approach based on the deep learning technique Convolutional Neural Network (CNN) and the particle swarm optimization algorithm (PSO) to optimize the filters in the convolutional layers, the second proposed method describes a classification approach conventional based on texture descriptors, using phylogenetic indices, the eXtreme Gradient Boosting (XGBoost) algorithm and the PSO algorithm to optimize the XGBoost hyperparameters, and finally, the third proposed method details a hybrid classification approach based on the CNN technique, the XGBoost algorithm and the PSO algorithm to optimize the type of connection used in the convolutional layers. The proposed methods were evaluated on the databases Prostate 3T and PROMISE12 using the performance metrics Dice similarity coefficient, relative volume difference, volumetric similarity, average distance surface and Hausdorff distance. The results of the application of the first method showed 87.67%, 2.83%, 0.96, 0.89 mm, and 13.65 mm, respectively, in the corresponding values of the mentioned performance metrics. The second proposed method obtained 85.64%, 7.68%, 0.96, 1.22 mm, and 15.13 mm, respectively. Finally, the proposed third method reached 87.65%, 3.18%, 0.96, 0.88 mm, and 13.51 mm, respectively. It was found that the first and the third method showed similar results in the segmentation of the prostate, being superior to the results obtained in the second method. In addition, the third method had a lower standard deviation in the metrics and a higher rate of hit in the prostate superpixels than the other methods. The experimental results demonstrate the performance potential of the proposed methods compared to those recently published in the literature.