Discrimination of ground glass opacity on lung HRCT images using visual complexity measurements

We investigated two image features based on visual complexity measurements: the fractal dimension (FD) and the Lempel-Ziv Complexity (LZC), and evaluated their performance in differentiating GGOs from normal areas on lung HRCT images. The database of this study contains 86 rectangular ROIs (44 Normal, 42 GGO) of 15*15 pixels. The features of FD and LCZ extracted from these ROIs were input to a linear classifier to predict their classification. When the two features were used individually, they respectively yielded areas under the ROC curve (AUC) of 0.837 and 0.902; 75.6%/80.2% of ROIs were correctly classified when training and testing in a re-substitution procedure; while 75.6%/79.1% were correctly classified when jackknifing was used. On condition that both features were input to the classifier, an AUC of 0.97 was achieved; meanwhile the overall accuracy increased up to 90.7%. The promising results demonstrated FD and LZC's potential in GGO discrimination.