Contributions to the design and implementation of algorithms for the classification of hyperspectral images of brain tumors in real-time during surgical procedures

• Autores: Himar Antonio Fabelo Gómez
• Directores de la Tesis: Gustavo Marrero Callicó (dir. tes.), Roberto Sarmiento (dir. tes.)
• Lectura: En la Universidad de Las Palmas de Gran Canaria ( España ) en 2019
• Idioma: español
• Tribunal Calificador de la Tesis: Antonio Núñez Ordóñez (presid.), Francisco Mario Hernández Tejera (secret.), J. Uceda (voc.), Fons van Der Sommen (voc.), Caifeng Shan (voc.)
• Programa de doctorado: Programa de Doctorado en Tecnologías de Telecomunicación e Ingeniería Computacional por la Universidad de Las Palmas de Gran Canaria
• Materias:
  • Ciencias tecnológicas
    • Tecnología electrónica
• Enlaces
  • Tesis en acceso abierto en: acceda
• Resumen

  • Hyperspectral imaging (HSI) allows the acquisition of large numbers of spectral bands throughout the electromagnetic spectrum with respect to the surface of scenes captured by sensors. Using this information and a set of complex classification algorithms, it is possible to determine which material or substance is located in each pixel. One of the major benefits of this technology is that it can be used as a guidance tool during brain tumor resections. Unlike other tumors, brain tumor infiltrates the surrounding normal brain tissue and thus their borders are indistinct and extremely difficult to identify to the surgeon’s naked eye. The surrounding normal brain tissue is critical and there is no redundancy, as in many other organs, where the tumor is commonly resected together with an ample surrounding block of normal tissue. In brain tumors, it is essential to accurately identify the margins of the tumor to resect as less healthy tissue as possible. In this sense, the work performed in this thesis aims to exploit the characteristics of HSI to develop an intraoperative demonstrator capable of performing a precise localization of malignant tumors during brain surgical procedures, expecting to improve the outcomes of the surgery. This demonstrator was able to generate thematic maps of the exposed brain surface using spectral information of the range comprised between 400 and 1000 nm, achieving intraoperative processing time (~1 min). These thematic maps distinguish between four different classes (normal tissue, tumor tissue, hypervascularized tissue and background), where the tumor boundaries can be easily identifiable. The thematic maps demonstrate that the system did not introduce false positives in the parenchymal area when no tumor was present and it was able to identify low-grade tumors that were not used to train the brain cancer detection algorithm, resulting in a robust and generalized classification algorithm. Additionally, a preliminary study to improve the accuracy results using deep learning techniques was performed. This study achieved promising results in the discrimination between tumor and normal brain tissue, being a suitable method to further improve the outcomes of the intraoperative demonstrator and, hence, the outcomes of the neurosurgical procedures.