Previsão da evolução de doença oncológica a partir da análise de imagens de PET scan

Abstract

An oncology patient response to a treatment can be assessed by the evaluation of PET/CT scans. One should notice this analysis is not valid to every oncology disease. This medical imaging technique is the combination of two other: PET (Positron Emission Tomography), that offers an image related to the metabolic activity, through the injection of radioactive isotopes; CT (Computed Tomography) provides an image related to the patient's morphology, with the help of x-rays. The goal of this work consists in the development of an approach to obtain models capable of predicting the outcome of the evolution of oncology diseases after its treatment. In this work, neuroendocrine tumors and Hodgkin lymphomas images were used, provided by the Institute of Oncology of Porto. The images were reduced to regions of interest, where the lesions were located, having been extracted features from PET/CT scans obtained before the treatment took place, creating the data set to be processed by Machine Learning techniques, alongside patients clinical data and treatment outcome. Several normalization techniques have been tested, with five supervised learning methods of different types, whose parameters were tuned. Two feature selection types were also used and techniques to deal with the class imbalance problem. The results obtained include predictions with f-scores of 0.3745, 0.7135, 0.4143 and 0.7705 for each of the treatment outcomes. The reduced number of instances combined with data distribution imbalance between the different outcome classes has maybe led to under-performance. Thus, to keep developing this work in the future, one step to take would be gathering more data from new oncology patients.