Tumor Subtype-Specific Parameter Optimization in a Hybrid Active Surface Model for Hepatic Tumor Segmentation of 3D Liver Ultrasonograms

Segmentation of hepatic tumors is a clinically demanding task for improving reliability in diagnosis and treatment procedures, and yet remains a challenging problem due to their highly noisy, low contrast, and blurry imaging nature. However, once correctly segmented, the shape and volume information of a tumor may provide useful information for radiological decision making. In this study, we propose an active surface model. The model combines edge, region, and contour smoothness energies. We extracted qualitative appearance features from three hepatic tumor subtypes and use them to adjust the weights of the energy terms in order to determine an optimized set of parameters for each tumor subtype. The performance of the developed method was evaluated with a dataset of 60 cases including 18 hepatic simple cysts, 18 hemangiomas, and 24 hepatocellular carcinomas, as determined by the radiologist's visual assessment. Evaluation of the results showed that our proposed method produced tumor boundaries that were equal to or better than acceptable in 87% of cases.