A New Two-Stage Hierarchical Framework for Mammogram Retrieval

In this work, we described a new two-stage hierarchical framework for mammogram retrieval. We tested the proposed approach on the reference library from USF-DDSM. For each query ROI (region of interest), the proposed scheme first computes its 14 texture and shape features, then the voting method based on five classifiers is used to classify the ROIs in the reference library, this phase eliminates ROIs that are very semantic different from the query ROI. Next, the scheme uses the k-nearest neighbor K-NN algorithm to select the 40 most ldquosimilarrdquo regions with the same computational level from a large reference library, and computes the mutual information (MI) between the query ROI and each of the selected ROIs from the database. Finally, 15 regions with the highest MI values, which seem to be the most visual similar to the query ROI, are selected and displayed to the observers. We tested the proposed method involving 80 queries, on average the 73.8% of the displayed ROIs are considered semantic and visual similar to the query region. Preliminary experimental results showed that the proposed scheme can achieve better accuracy than the single-classifier based method and the vector distance based similarity metric.