An Efficient Method for PACS Medical Image Retrieval

Content-based image retrieval (CBIR) enables the identification of visually similar images from large image repositories. Conventional CBIR methods often represent each image using a global feature vector, which may fail to preserve spatially localized patterns and structural information. This paper proposes a region-wise multi-feature image retrieval framework that partitions each image into a 4×4 grid of non-overlapping sub-regions and extracts complementary edge, intensity/color, and texture features from each region. Rather than relying on fixed spatial correspondences or a single global feature vector, the proposed framework models the correspondence between the sub-regions of a query image and those of a candidate image as a weighted bipartite graph. The resulting distance matrix is used by the standard Hungarian algorithm to determine the minimum-cost perfect matching between image regions. The proposed framework is quantitatively evaluated on the WANG/Corel-1K benchmark against representative color-, texture-, and graph-based CBIR methods. The results show a mean retrieval precision of 69% under the reported feature-weight configuration and a precision of 59% at 100% recall, compared with 55%, 46%, and 43% for the baseline methods, respectively. In addition, a qualitative case study using clinical femoral X-ray images is presented to illustrate the potential application of the framework to medical image retrieval. The findings demonstrate the retrieval performance of the proposed approach on the general-purpose benchmark, while further quantitative evaluation on dedicated medical image datasets is required to establish its effectiveness in medical imaging applications.