Autodetection of semantic features in pulmonary nodules
This study seeks to use the large number of cases available in the NLST to investigate that such features can be auto-detected with high throughput unsupervised segmentation approaches of both normal structures and nodules to match module semantics (e.g. NCCN ratings) assigned by expert radiologists. The study further seeks to establish whether such semantics can be correlated with tumor growth in small indeterminate nodules. The normal structures obtained from high throughput segmentation of this study will be contributed to the archive as RT structure sets.
a) Demonstrate that normal structures of the lung (lung cavities, trachea, vessel tree, esohpagus, spinal cord) etc. can be auto segmented with high quality and used to set boundaries on lung nodule segmentation and assign semantic features such as lobe location, degree of attachment, vessel penetration, etc.
b) Demonstrate the extent to which these features match manual ratings of expert radiologists using ontologies such as LungRads and NCCN ratings.
c) Using longitudinal data from the NLST, demonstrate that both sets of semantic features can be correlated with tumor growth and are invariant for different acquisition kernels of the CT data sets.
d) Contribute to the NLST archive normal structures obtained from the study as RT structure sets. These can be used for other investigators for nodule segmentation and characterization.
David Gering, Ph.D.
Hao Wang, Ph.D.