Bayesian Saliency Model for Focal Liver Lesion Enhancement and Detection

Focal liver lesion enhancement and detection has an essential role for the computer-aided diagnosis and characterization of lesion regions in CT volume data. This paper proposes a novel focal lesion enhancement strategy by extracting a lesion saliency map, which represents the deviation degree of the uncommon or lesion tissue from the common tissues (liver and vessel) in CT volumes. The saliency map can be constructed by exploring the existing probability of lesion for any voxel. However, due to the large diversity of liver lesions, it is difficult to construct an universal model for all types of lesions. Therefore, this study proposes to construct probability models of the common tissues, which have comparably small variability even for different samples and is relatively easy to obtain the prototype regions even from the under-studying CT volume. In order to robustly and flexibly characterize the common tissues, we explore a Bayesian framework by combining a general model, which is constructed oriented to all CT samples, and an adaptive model, which is constructed specific to the under-studying CT sample, for calculating the existing probability of the common tissues (liver or vessel). Then, the saliency map (the existing probability) of focal lesion can be deduced from that of liver or vessel. The advantages of our proposed strategy mainly include three aspects: (1) it only needs to prepare the prototypes of common tissue such as liver or vessel region, which are easily obtained in any CT liver volume; (2) it proposes to combine the general and adaptive model as Bayesian framework for more robust and flexible characterization of the common tissue; (3) dispensable to remove the other different structure such as vessel in liver volume as a pre-processing step. Experiments validate that the proposed Bayesian-based saliency model for focal liver lesion enhancement can perform much better than the conventional approaches such as EM, EM/MPM based lesion detection and segmentation methods.