Evaluation of Local Texture Features in Content-based Medical Image Retrieval Task

Content-Based Medical Image Retrieval (CBMIR) is an effective tool for managing extensive collections of medical images. A CBMIR system can be implemented by utilizing hand-crafted local texture features stored in a database, enabling the retrieval of relevant images based on similarity matching. While various local texture descriptors are available for describing medical images, their performance in CBMIR tasks has not been thoroughly evaluated. This study develops a Query-by-Example (QBE) CBMIR system and evaluates the performance of 25 local texture descriptors in CBMIR task. The evaluation uses three publicly available datasets: NEMA-CT, OASIS-MRI, and Emphysema-CT. Four metrics—Average Retrieval Precision (ARP), Average Retrieval Recall (ARR), F1-score, and Average Normalized Modified Retrieval Rank (ANMRR) are employed for performance assessment. The results reveal that on the NEMA-CT dataset, the Local Directional Gradient Pattern (LDGP) texture descriptor with 4 × 4 sub-regions using Manhattan distance performs best. For the OASIS-MRI and Emphysema-CT datasets, the Threshold Local Binary AND Pattern (TLBAP) and Local Adjacent Neighborhood Average Difference Pattern (LANADP) yield superior results. This work provides valuable insights for researchers selecting local texture descriptors for CBMIR tasks.