Evaluation of 3D anatomical surfaces indexing for surgery planning

Finding pertinent images in large picture archiving systems, for advanced medical practice support is becoming increasingly difficult. One possible solution to such emerging problem is image indexing. This work proposes to evaluate the indexing and retrieval performance of various 3D anatomical indexing approaches, in order to assist surgery planning based on similar cases. The evaluation examines the indexing performance of 5 feature descriptors (simple statistic, cord- based, shape distribution, surface curvature, and 3D Hough transform) and the retrieval performance of 5 similarity measures (the Minkowski norms L1 L2 and L(infinity), the Bhattacharyya distance, and the chi2-divergence). A database of 21 patients, with an average of 11 3D anatomical surfaces per patient was used. The combined performance of feature descriptors and similarity measurements was evaluated with the Bull-Eye Percentage score. Experimental results indicate that there are several possible optimal indexing and retrieval approaches, depending on the surface characteristics.