Learning lung nodule similarity using a genetic algorithm

The effectiveness and efficiency of content-based image retrieval (CBIR) can be improved by determining an optimal combination of image features to use in determining similarity between images. This combination of features can be optimized using a genetic algorithm (GA). Although several studies have used genetic algorithms to refine image features and similarity measures in CBIR, the present study is the first to apply these techniques to medical image retrieval. By implementing a GA to test different combinations of image features for pulmonary nodules in CT scans, the set of image features was reduced to 29 features from a total of 63 extracted features. The performance of the CBIR system was assessed by calculating the average precision across all query nodules. The precision values obtained using the GA-reduced set of features were significantly higher than those found using all 63 image features. Using radiologist-annotated malignancy ratings as ground truth resulted in an average precision of 85.95% after 3 images retrieved per query nodule when using the feature set identified by the GA. Using computer-predicted malignancy ratings as ground truth resulted in an average precision of 86.91% after 3 images retrieved. The results suggest that in the absence of radiologist semantic ratings, using computer-predicted malignancy as ground truth is a valid substitute given the closeness of the two precision values.