Feature based differentiation of benign, malignant and atypical neurofibroma in FDG-PET scans

Objectives The aim of this study is to investigate whether information contained in Gray-Level Co-occurrence Matrix (GLCM) texture features can provide information supplemental to SUV in distinguishing between benign, malignant, and atypical neurofibromas in FDG PET scans.

Methods Scans of 49 patients diagnosed with neurofibroma were analyzed. A total of 63 lesions were identified by a radiologist. These comprised 36 benign, 14 malignant and 11 atypical, on the basis of histology and clinical follow-up. Images were reconstructed with 3D OSEM. The GCLM was used to calculate texture features at each tumour location within a 5x5x5 voxel neighbourhood around the maximum voxel. Background was excluded by a simple heuristic threshold. Mean and standard deviations (SD) were computed over all directions for each texture feature, yielding a total of 12 features, plus local features including SUVmax and SUVmean. In order to define the co-occurrence matrix, the maximal SUV range (0-16 here) is divided into n equal-width bins. Five binnings were explored from n=8 to n=128. Intermediate ranges (n=32,64) were found to be optimal.

Results Texture SD of energy, entropy and inverse difference moment (IDM) were found to contain the maximum complementary information relative to SUV (correlation coefficient = 0.3). This result was found to hold independent of the sub-category of the tumour population, and also when all categories were considered as a whole. A line in the 2-D space defined by IDM SD and SUVmean was seen to classify malignant and benign with fewer false positives than SUV alone.

Conclusions Texture features provide information complementary to SUVmax. Correlating this additional information with histopathology/clinical outcome may provide useful knowledge about the link between higher order information and disease severity/staging. This is important in the case of neurofibroma with intermediate SUV values (2-3.5) where SUV alone is equivocal.

Research Support This work was funded in part by the KCL/UCL CRUK/EPSRC Comprehensive Cancer Imaging Centre