Lung scintigraphy clustering by texture analysis.

The efficiency of texture analysis parameters, describing the organization of grey level variations of an image, was studied for lung scintigraphic data classification. Twenty one patients received a 99mTc-MAA perfusion scan and 81mKr and 127Xe ventilation scans. Scans were scaled to 64 grey levels and 100 k events for inter subject comparison. The texture index was the average of the absolute difference between a pixel and its neighbors. Energy, entropy, correlation, local homogeneity and inertia were computed using co-occurrence matrices. A principal component analysis was carried out on each parameter for each type of scan and the first principal components were selected as clustering indices. Validation was achieved by simulating 2 series of 20 increasingly heterogeneous perfusion and ventilation scans. For most of the texture parameters, one principal component could summarize the patients data since it corresponded to the relative variances of 67%-88% for perfusion scans, 53%-99% for 81mKr scans and 38%-97% for 127Xe scans. The simulated series demonstrated a linear relationship between the heterogeneity and the first principal component for texture index, energy, entropy and inertia. This was not the case for correlation and local homogeneity. We conclude that heterogeneity of lung scans may be quantified by texture analysis. The texture index is the easiest to compute and provides the most efficient results for clinical purpose.