Indexed in

External link

License and use

Citations

Altmetrics

Analysis of institutional authors

Sanz Alvaro, CesarAuthorJuarez Martinez, EduardoAuthorVazquez Valle, GuillermoCorresponding AuthorMartin Perez, AlbertoCorresponding AuthorVilla Romero, ManuelCorresponding AuthorSancho Aragon, JaimeCorresponding AuthorRosa Olmeda, GonzaloCorresponding AuthorCebrian Castel, Pedro LuisCorresponding AuthorSutradhar, PallabCorresponding AuthorMartinez De Ternero Ruiz, AlejandroCorresponding AuthorChavarrias Lapastora, MiguelAuthor

Share

Sparse to Dense Ground Truth Pre-Processing in Hyperspectral Imaging for In-Vivo Brain Tumour Detection

Publicated to:2023 Ieee International Conference On Metrology For Extended Reality, Artificial Intelligence And Neural Engineering, Metroxraine 2023 - Proceedings. 272-277 - 2023-01-01 (), DOI: 10.1109/MetroXRAINE58569.2023.10405811

Authors: Vazquez G; Martín-Pérez A; Villa M; Sancho J; Rosa G; Cebrián PL; Sutradhar P; De Ternero AM; Perez-Nuñez A; Jimenez-Roldan L; Lagares A; Chavarrías M; Juarez E; Sanz C

Affiliations

Abstract

Image segmentation tasks often require fully annotated datasets where the boundaries of the elements to be identified appear accurately marked. However, such detailed ground truth is hard to obtain mainly because it usually involves a time consuming procedure. In biomedical applications, this may imply that the medical specialist in charge of the labelling process can only mark a few sparse samples belonging to the principal elements of interest. In a context of in-vivo brain tumour detection through machine learning techniques and hyperspectral imaging, such sparse ground truth restricts the training of the classifiers to work at pixel level. In addition, the absence of a dense region localising the tumour makes it more difficult to assess the quality of the segmentation with objective metrics. To address these problems, two ground truth pre-processing methodologies are proposed in order to obtain a dense ground truth map of the tumour region from sparse annotations: a BFS (Breadth-First Search)-based method and an adaptation of the SLIC superpixels algorithm. The proposed work is tested by analysing the effect it has on the training of a convolutional neural network with an autoencoder-type architecture. The results are validated comparing the training metrics obtained using a sparse ground truth, the proposed methodology and the state of the art techniques.

Keywords