Abstract
We present a new method for the automatic generation of patient-specific, feature-based multi-dimensional transfer functions used in the visualization of liver blood vessels and tumors in CT datasets. The method automatically extracts the geometrical structure of the vessels and tumors with a multi-scale eigenanalysis of the image Hessian matrix. It then uses this information to optimize the transfer function based on energy minimization in a variational framework. The method overcomes key drawbacks of existing volume visualization techniques, which are limited to predefined transfer functions, require lengthy manual adjustment based on CT iso-values of the structure of interest, and often produce suboptimal results. We applied the method to five clinical data sets with 72-110 slices, obtained transfer functions for each in 53-137 seconds, and produced real-time visualizations. The visualizations were evaluated and compared to those obtained by two existing methods by an expert radiologist, who deemed them superior in detail and discriminating power.
Keywords: CT volume visualization, liver,
vessels, tumors, transfer function, variational principle