UNCERTAINTY QUANTIFICATION OF SEGMENTATIONS OF ORGANS AT RISK USING A PARTIALLY LABELED DATASET IN THE CONTEXT OF PROTON THERAPY

Abstract

Cancer has been a group of illness that has attacked humanity throughout history. For that, medical and physical treatments have been developed and, with them, radiation therapies. They are based on the destruction of those cancerous cells by means of radioactivity. As a branch, proton therapy has led the state-of-the-art due to its precision and improvement in patient¿s prognosis. Nonetheless, that accuracy has risen worries regarding targeting the lesions. Because of that, planning is the most important step of the whole treatment for a proper outcome. One of the steps is the delimitation of organs at risks that, hegemonically, has been performed manually by qualified professionals. When considering those volumes, borders containing the tumors are selected, including healthy tissues to prevent overlooking cancer cells. That irradiation of healthy organs may cause deterioration of said structures. That technique was quickly complemented with artificial intelligence, allowing automatic segmentations of organs at risk. This field of study has been present in literature for decades, looking for an ideal approach to tackle this particular issue. The state-of-the-art has evolved up to the invention of a model based on the use of adaptive simple architectures, the nnUNet. Nonetheless, despite the presence of deep learning models that are able to contour both organs and tumors, surrounding tissues are still irradiated to prevent reconcurrence of the illness. With an uncertainty quantification study regarding the predictor model, the possibility of reducing tumoral volumes and minimizing those healthy tissues from being irradiated is considered. Motivated by this, this Bachelor¿s thesis has developed a prediction based on the nnUNet with the help of Centro de Protonterapia Quironsalud. A partially labeled dataset was provided for model training, containing 250 cases of Head and Neck cancer. A first evaluation was performed on the predictions of the model using metrics like Dice Coefficient (resulting in ¿ 0.73) and Hausdorff Distance (being ¿ 3.312mm), returning promising results that led into the study of uncertainty quantification of this model. In it, small variations to the volumes were performed to quantify the variability of the results. That study entailed statistical analysis of the distances between segmentations generated by the same patient with slight changes. It turned out to be robust, highlighting the lack of sense of using complex Deep Learning models instead of simple architectures that are able to adapt to the environment they are being used in, allowing real clinical implementation. Those results and its interpretability allowed the creation of a tool where the whole process is encapsulated.