Latent Semantic Association for Medical Image Retrieval

Abstract

In this work, we propose a Latent Semantic Association Retrieval(LSAR) method to break the bottleneck of the low-level feature based medical image retrieval. The method constructs the high-level semantic correlations among patients based on the low-level feature set extracted from the images. Specifically, a Pair-LDA model is firstly designed to refine the topic generation process of traditional Latent Dirichlet Allocation (LDA), by generating the topics in a pair-wise context. Then, the latent association, called CCA-Correlation, is extracted to capture the correlations among the images in the Pair-LDA topic space based on Canonical Correlation Analysis (CCA). Finally, we calculate the similarity between images using the derived CCA-Correlation model and apply it to medical image retrieval. To evaluate the effectiveness of our method, we conduct the retrieval experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) baseline cohort with 331 subjects, and our method achieves good improvement compared to the state-of-the-art medical image retrieval methods. LSAR is independent on problem domain, thus can be generally applicable to other medical or general image analysis.