Semantic modeling for multi-level medical image semantics retrieval

In this study, a multi-level medical image semantic modeling approach based on fuzzy Bayesian networks is proposed. Its two forms are built. The one is a Bayesian network embedding Conditional Gaussian (CG) models, called BN-CG, and another is a Bayesian network embedding Gaussian mixture model (GMM), called BN-GMM. CG and GMM are employed to implement a fuzzy procedure to perform the soft quantification of the continuous visual feature of the medical images, which extract the middle level semantics of the pathological objects, using the probability as the confidence score. Finally, a Bayesian network is utilized to combine these middle level semantics to build a multi-level semantic model. BN-CG and BN-GMM model are tested at multiple levels of semantics by applying a small set of astrocytona MRI (Magnetic Resonance Imaging) image samples. The experiment results show that this approach is very effective to enable the auto-annotation and interpretation of astrocytona MRI images. These models outperform the Bayesian network-based crisp quantification model using k-nearest neighbor classifiers (K-NN). This study provides a novel way to assist radiologist to retrieve medical images.