Modelo para a classificação de nódulos pulmonares pequenos usando descritores radiomics

• Ano de defesa: 2017
• Autor(a) principal: Lima Filho, Ailton Felix de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alagoas; Brasil; Programa de Pós-Graduação em Informática; UFAL
• Palavras-chave em Português: Tecnologia da informação; Informática médica; Diagnóstico auxiliado por computador; Nódulo pulmonar; Câncer de pulmão; Diagnóstico por imagem; Radiomics; Information technology; Medical informatics; Diagnostics aided computer; Lung node; Lung cancer; Diagnosis by image; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://www.repositorio.ufal.br/handle/riufal/2000

Lung cancer is a disease characterized as abnormal cells growth that invade and destroy neighboring tissues, accounting for many deaths around the world. An early diagnosis, usually performed based on qualitative information extracted from CT images, brings greater chances of cure and treatment options for the patient, however, due to the challenges in the medical image interpretation process, mainly for small pulmonary nodules (<10mm), the diagnosis becomes clinically difficult, making the clinical decision complex. Due to the variability and complexity of the diagnosis of small pulmonary nodules, Computer-Aided Diagnosis (CAD) tools based on image features, provides assistance to the radiologist in order to achieve a better accuracy of nodule classification (probable malignant or benign), by acting as a second opinion to the specialist. The use of radiomics features allows a quantitative diagnosis when compared to the recent qualitative strategies of cancer evaluation, significantly reducing the problem of variability in diagnosis. However, discovering the relevant content/features is still a necessity in order to improve the CAD systems performances. The aim of this study was to develop a classification model for small pulmonary nodules using radiomics features extracted from the nodule microenvironment. It was also evaluated the hypotheses test that considering the parenchyma region around the nodule allows an improvement in the small pulmonary nodules classification. The developed classification model obtained the best Area Under the ROC curve (AUC) of 0.875 ± 0.048 with the Multilayer Perceptron (MLP) algorithm with a 10-fold cross-validation in the classification of 214 pulmonary nodules with diameters between 5 and 10mm. The results showed the relevance of radiomics features for the classification of small pulmonary nodules. The use of the pulmonary parenchyma region improved the model performance, proving the hypothesis test. The nodules classification is a challenging area for specialists due to the natural complexity of diagnosis lesions, however, critical for patient survival diagnosed with cancer. Therefore, advances in this area are extremely important.