Visual Analysis of Lung Cancer Imaging Data Using Multidimensional Projections Techniques
I. Tolios · Research Repository (Delft University of Technology) · 2016
Lung tumors are heterogeneous entities consisting of distinct intra-tumor regions with different biological characteristics. Currently, the only means to explore and analyze these distinct intra-tumor regions is to inspect several image acquisitions. From these, per-voxel functional tissue parameters, representative of tissue properties, are extracted. This results in a high-dimensional imaging-derived feature space. However, the exploration and analysis of tumor heterogeneity based on the derived imaging data is a complex and time demanding task, due to its multi-modality and high dimensionality. Visual Analytics techniques can help optimize this process. Views based on dimensionality reduction are a common strategy to offer insight in high-dimensional feature spaces by constructing low-dimensional representations of the initial high-dimensional data. There is a great range of dimensionality reduction techniques. For this reason, we have to find the most suitable for visualizing and analyzing lung cancer imaging data. To this end, we selected and evaluated those who according to our opinion meet our visualization requirement. This requirement is the preservation of the local structure of the data in the low-dimensional representation, as an indication of intra-tumor regions being grouped together in the high-dimensional space. For the selection of the examined techniques, we evaluate them both qualitatively and quantitatively. The qualitative evaluation is performed by visually inspecting the low-dimensional mappings, while the quantitative by applying specific metrics. The best performing techniques are introduced in an existing visualization framework [1] and their anatomical coherence is examined. However, additional functionalities are introduced, hereby enhancing its exploratory value. Side-by-side comparison of multiple low-dimensional projections and automatic clustering on the embeddings are added in the tool. The proposed tool is evaluated by our collaborating clinical researcher with three real patient cases. Overall, the tool provided promising results, making easier the identification, inspection and hypothesis generation for intra-tumor regions.