Intensity overlap often occurs in medical images, making it difficult to identify different anatomical structures using
intensity alone. Research studies have shown that texture is an important component in quantifying the visual
appearance of anatomical structures, and is therefore valuable in the analysis, interpretation, and retrieval of lung
The goal of our research study is to present a comparison between the different texture models: Gabor filters, Markov
Random Field (MRF), and global & local co-occurrence. For comparison purposes we utilized Manhattan, Euclidean,
and Chebyshev distances for one-dimensional feature vectors (global co-occurrence) while for two-dimensional feature
comparison (local co-occurrence, Gabor filters, and MRF) we utilized the similarity measures Chi-Square and Jeffrey-
Divergence. Local co-occurrence contains many different variable aspects in its design that can considerably change the
success of its results. A thorough examination of local co-occurrence's variables is discussed.
All of the discussed texture models are presented in the context of our previous Content-Based Image Retrieval (CBIR)
System . BRISC utilizes the Lung Image Database Consortium (LIDC) database. We have found that Gabor and
MRF texture descriptors produce the best retrieval results regardless of the nodule size, number of retrieved items or
similarity metric with an average precision of 88%. Global co-occurrence performed the worse at 44% precision yet
when co-occurrence was performed locally (local co-occurrence) the precision results improved to 64%. A combination
of all the features worked the best with 91% precision.