- -

Estudio de técnicas de deep learning para una segmentación automática de imágenes de resonancia magnética de metástasis óseas

RiuNet: Institutional repository of the Polithecnic University of Valencia

Share/Send to

Cited by

Statistics

Estudio de técnicas de deep learning para una segmentación automática de imágenes de resonancia magnética de metástasis óseas

Show simple item record

Files in this item

dc.contributor.advisor Moratal Pérez, David es_ES
dc.contributor.author Souto Pastor, Marina es_ES
dc.date.accessioned 2020-04-13T23:08:39Z
dc.date.available 2020-04-13T23:08:39Z
dc.date.created 2019-09-30
dc.date.issued 2020-04-14 es_ES
dc.identifier.uri http://hdl.handle.net/10251/140617
dc.description.abstract [EN] Bone metastases are a common complication in some high incidence types of cancer, like prostate or breast cancer. The complications associated with bone metastases include bone pain, fractures and spinal cord compression. Most part of bone metastases are irreversible and treatments are focused on slowing the growth of the lesions. In the United States, 17% of the total direct medical cost was employed treating bone metastases. In order to improve the health of the patients and cut down medical costs, early detection is crucial. Some studies have shown that Whole-Body MRI has the potential to become the best method for diagnosis but there are still some difficulties left. One patient can have multiple bone metastases all over the skeleton in different sizes. This makes diagnosing bone metastases a tough task for the radiologists and because of the irregular shapes of the bone metastases, changes in size are also difficult to measure. The goal of this project is to provide an automatic tool for the segmentation of bone metastasis, making it easier for the clinicians to diagnose and to control the size of the present metastases. Using different modalities of MRI (T1 and B1000) and different patch sizes (16x16x16 and 32x32x32) a convolutional neural network (UNet) was trained. The segmentations predicted by each U-Net employing one modality and size, were later combined into one final segmentation. The best results achieved with this approach are the following: a correct detection of 37 bone metastases out a total of 100 with 67 false positives using k fold cross-validation and a dataset of 6 different patients with multiple acquisitions making a total of 100 lesions. es_ES
dc.description.abstract [ES] En este trabajo se investiga la capacidad de las redes neuronales, en concreto redes neuronales convolucionales (CNN), como una herramienta para la segmentación automática de metástasis óseas locales en imágenes multimodales de resonancia magnética. es_ES
dc.language Inglés es_ES
dc.publisher Universitat Politècnica de València es_ES
dc.rights Reconocimiento (by) es_ES
dc.subject Segmentación automática es_ES
dc.subject 3D CNN es_ES
dc.subject Metástasis óseas es_ES
dc.subject U-Net es_ES
dc.subject MRI es_ES
dc.subject Automatic segmentation es_ES
dc.subject Bone metastasis es_ES
dc.subject.classification TECNOLOGIA ELECTRONICA es_ES
dc.subject.other Máster Universitario en Ingeniería Biomédica-Màster Universitari en Enginyeria Biomèdica es_ES
dc.title Estudio de técnicas de deep learning para una segmentación automática de imágenes de resonancia magnética de metástasis óseas es_ES
dc.type Tesis de máster es_ES
dc.rights.accessRights Abierto es_ES
dc.contributor.affiliation Universitat Politècnica de València. Departamento de Ingeniería Electrónica - Departament d'Enginyeria Electrònica es_ES
dc.contributor.affiliation Universitat Politècnica de València. Escuela Técnica Superior de Ingenieros Industriales - Escola Tècnica Superior d'Enginyers Industrials es_ES
dc.description.bibliographicCitation Souto Pastor, M. (2019). Estudio de técnicas de deep learning para una segmentación automática de imágenes de resonancia magnética de metástasis óseas. http://hdl.handle.net/10251/140617 es_ES
dc.description.accrualMethod TFGM es_ES
dc.relation.pasarela TFGM\114701 es_ES


This item appears in the following Collection(s)

Show simple item record