Case-based fracture image retrieval.

PURPOSE: Case-based fracture image retrieval can assist surgeons in decisions regarding new cases by supplying visually similar past cases. This tool may guide fracture fixation and management through comparison of long-term outcomes in similar cases.
METHODS: A fracture image database collected over 10 years at the orthopedic service of the University Hospitals of Geneva was used. This database contains 2,690 fracture cases associated with 43 classes (based on the AO/OTA classification). A case-based retrieval engine was developed and evaluated using retrieval precision as a performance metric. Only cases in the same class as the query case are considered as relevant. The scale-invariant feature transform (SIFT) is used for image analysis. Performance evaluation was computed in terms of mean average precision (MAP) and early precision (P10, P30). Retrieval results produced with the GNU image finding tool (GIFT) were used as a baseline. Two sampling strategies were evaluated. One used a dense 40 × 40 pixel grid sampling, and the second one used the standard SIFT features. Based on dense pixel grid sampling, three unsupervised feature selection strategies were introduced to further improve retrieval performance. With dense pixel grid sampling, the image is divided into 1,600 (40 × 40) square blocks. The goal is to emphasize the salient regions (blocks) and ignore irrelevant regions. Regions are considered as important when a high variance of the visual features is found. The first strategy is to calculate the variance of all descriptors on the global database. The second strategy is to calculate the variance of all descriptors for each case. A third strategy is to perform a thumbnail image clustering in a first step and then to calculate the variance for each cluster. Finally, a fusion between a SIFT-based system and GIFT is performed.
RESULTS: A first comparison on the selection of sampling strategies using SIFT features shows that dense sampling using a pixel grid (MAP = 0.18) outperformed the SIFT detector-based sampling approach (MAP = 0.10). In a second step, three unsupervised feature selection strategies were evaluated. A grid parameter search is applied to optimize parameters for feature selection and clustering. Results show that using half of the regions (700 or 800) obtains the best performance for all three strategies. Increasing the number of clusters in clustering can also improve the retrieval performance. The SIFT descriptor variance in each case gave the best indication of saliency for the regions (MAP = 0.23), better than the other two strategies (MAP = 0.20 and 0.21). Combining GIFT (MAP = 0.23) and the best SIFT strategy (MAP = 0.23) produced significantly better results (MAP = 0.27) than each system alone.
CONCLUSIONS: A case-based fracture retrieval engine was developed and is available for online demonstration. SIFT is used to extract local features, and three feature selection strategies were introduced and evaluated. A baseline using the GIFT system was used to evaluate the salient point-based approaches. Without supervised learning, SIFT-based systems with optimized parameters slightly outperformed the GIFT system. A fusion of the two approaches shows that the information contained in the two approaches is complementary. Supervised learning on the feature space is foreseen as the next step of this study.