Proceedings Article | 19 March 2008
KEYWORDS: Lung, Opacity, Classification systems, Emphysema, Radiology, Statistical analysis, Shape analysis, Machine learning, Medical imaging, Feature extraction
To find optimal binning, variable binning size linear binning (LB) and non-linear binning (NLB) methods were tested. In
case of small binning size (Q ≤ 10), NLB shows significant better accuracy than the LB. K-means NLB (Q = 26) is
statistically significant better than every LB. To find optimal binning method and ROI size of the automatic classification
system for differentiation between diffuse infiltrative lung diseases on the basis of textural analysis at HRCT
Six-hundred circular regions of interest (ROI) with 10, 20, and 30 pixel diameter, comprising of each 100 ROIs
representing six regional disease patterns (normal, NL; ground-glass opacity, GGO; reticular opacity, RO;
honeycombing, HC; emphysema, EMPH; and consolidation, CONS) were marked by an experienced radiologist from
HRCT images. Histogram (mean) and co-occurrence matrix (mean and SD of angular second moment, contrast,
correlation, entropy, and inverse difference momentum) features were employed to test binning and ROI effects. To find
optimal binning, variable binning size LB (bin size Q: 4~30, 32, 64, 128, 144, 196, 256, 384) and NLB (Q: 4~30)
methods (K-means, and Fuzzy C-means clustering) were tested. For automated classification, a SVM classifier was
implemented. To assess cross-validation of the system, a five-folding method was used. Each test was repeatedly
performed twenty times. Overall accuracies with every combination of variable ROIs, and binning sizes were statistically
compared.
In case of small binning size (Q ≤ 10), NLB shows significant better accuracy than the LB. K-means NLB (Q = 26) is
statistically significant better than every LB. In case of 30x30 ROI size and most of binning size, the K-means method
showed better than other NLB and LB methods. When optimal binning and other parameters were set, overall sensitivity
of the classifier was 92.85%. The sensitivity and specificity of the system for each class were as follows: NL, 95%,
97.9%; GGO, 80%, 98.9%; RO 85%, 96.9%; HC, 94.7%, 97%; EMPH, 100%, 100%; and CONS, 100%, 100%,
respectively.
We determined the optimal binning method and ROI size of the automatic classification system for differentiation
between diffuse infiltrative lung diseases on the basis of texture features at HRCT.
