Discriminative Random Field Modeling of Lung Tumors in CT Scans

The ability to conduct high-quality automatic 3D segmentation of tumors in CT scans is of high value to busy radiologists. Discriminative random fields (DRFs) were used to segment 3D volumes of lung tumors in CT scan data. Optimal parameters for the DRF inference were first calculated using gradient ascent. These parameters were then used to solve the inference problem using the graph cuts algorithm. Results of the segmentation were varied, with DRFs performing better on isolated tumors, but exhibiting bleed-through to adjacent tissues with similar intensities. Improvements can be made in the selection of features, discriminative models, and parameter optimization algorithm.