Texture analysis on ¹⁸F-FDG PET/CT images to differentiate malignant and benign bone and soft-tissue lesions

Abstract

Objective

The purpose is to develop and evaluate the ability of the computer-aided diagnosis (CAD) methods that apply texture analysis and pattern classification to differentiate malignant and benign bone and soft-tissue lesions on 18F-fluorodeoxy-glucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) images.

Methods

Subjects were 103 patients with 59 malignant and 44 benign bone and soft tissue lesions larger than 25 mm in diameter. Variable texture parameters of standardized uptake values (SUV) and CT Hounsfield unit values were three-dimensionally calculated in lesional volumes-of-interest segmented on PET/CT images. After selection of a subset of the most optimal texture parameters, a support vector machine classifier was used to automatically differentiate malignant and benign lesions. We developed three kinds of CAD method. Two of them utilized only texture parameters calculated on either CT or PET images, and the other one adopted the combined PET and CT texture parameters. Their abilities of differential diagnosis were compared with the SUV method with an optimal cut-off value of the maximum SUV.

Results

The CAD methods utilizing only optimal PET (or CT) texture parameters showed sensitivity of 83.05 % (81.35 %), specificity of 63.63 % (61.36 %), and accuracy of 74.76 % (72.82 %). Although the ability of differential diagnosis by PET or CT texture analysis alone was not significantly different from the SUV method whose sensitivity, specificity, and accuracy were 64.41, 61.36, and 63.11 % (the optimal cut-off SUV_max was 5.4 ± 0.9 in the 10-fold cross-validation test), the CAD method with the combined PET and CT optimal texture parameters (PET: entropy and coarseness, CT: entropy and correlation) exhibited significantly better performance compared with the SUV method (p = 0.0008), showing a sensitivity of 86.44 %, specificity of 77.27 %, and accuracy of 82.52 %.

Conclusions

The present CAD method using texture analysis to analyze the distribution/heterogeneity of SUV and CT values for malignant and benign bone and soft-tissue lesions improved the differential diagnosis on ¹⁸F-FDG PET/CT images.

Appendix

The texture is defined as the spatial interrelationship between and/or spatial arrangement of FDG uptake (or HU values) inside target lesions. Texture analysis is a set of mathematical methods that extract such information by some quantified parameters. We utilized three statistics-based texture analysis methods, which were the gray-level histogram analysis (GLHA) based method, the spatial gray-level dependence matrix (SGLDM) based method [14, 15] and the neighborhood gray-tone difference matrices (NGTDM) based method [16].

GLHA parameters

GLHA parameters include 4 statistics (mean, standard deviation, skewness, and kurtosis) calculated from gray-level distributions (histogram) of voxels. Since these parameters are directly derived from a histogram (Fig. 2a), they can only reveal how the gray levels of voxels statistically vary without considering spatial relationships.

SGLDM parameters

SGLDM parameters are calculated from co-occurrence matrix describing how often a voxel of gray-level i finds a neighbor voxel with gray-level j according to a certain spatial relationship (Fig. 2b), so they are related to spatial relationships of a pair of consecutive voxels [14]. There are 13 co-occurrence matrices in 3D by considering 13 different angular directions (26-connexity) and 1-distance relationship between consecutive voxels [15]. On these matrices, ten parameters (Table 2) characterizing heterogeneity are calculated by averaging values, respectively, derived from these 13 matrices. Equations 1 and 2 define entropy and correlation, respectively.

$${\text{Entropy }} = - \sum\limits_{i,j} {M_{1} \left( {i,\,j} \right) \times \log M_{1} \left( {i,\,j} \right)}$$

(1)

$${\text{Correlation}} = \frac{1}{{\sigma_{x} \sigma_{y} }}\sum\limits_{i} {\sum\limits_{j} {(i - u_{x} )(j - u_{y} )M_{1} (i,\,j)} } ,$$

(2)

where M ₁(i, j) is a co-occurrence matrix, u _x , u _y , σ _x, and σ _y are the means and standard derivations of M _x  = ∑ _j  M ₁(i, j) and M _y  = ∑ _i  M ₁(i, j), respectively. The definitions of the other eight parameters can be referred to in the literature [14, 15].

NGTDM parameters

NGTDM parameters contain information about how the gray level of a voxel varies from all its neighbors (Fig. 2c). They are calculated from neighborhood gray-tone difference matrix M ₂(i), whose entry describes the differences between voxels having gray-level i and their neighbors (26 neighbors in 3D). Five NGTDM parameters (Table 1) are calculated from the neighborhood gray-tone difference matrix. For example, Eq. 3 gives the definition of coarseness.

$${\text{Coarseness}} = \left[ {\upvarepsilon + \sum\limits_{i} {M_{2} \left( i \right) \times p\left( i \right)^{ - 1} } } \right] ,$$

(3)

where p(i) is the probability of gray-level i existing inside the lesion, and ɛ is a small number to prevent coarseness becoming infinite. The definitions of the other four NGTDM can be referred to in the literature [16].

The units of texture parameters

All of the 19 parameters are statistical measures, and their units can be derived from the equations of their definitions. According to these definitions, each parameter is calculated from two kinds of term that are related to the intensity and probability of the intensity. Because the probability is a dimensionless measure, the units of the texture parameters are derived only from the voxel intensity. The SUV value is a dimensionless measure; however, the CT value is expressed in Hounsfield units (HU). Therefore, the same texture parameter that is calculated on PET or CT images has different units.

Since the SUV value does not have a unit, none of the texture parameters that are calculated on PET images have units either. For texture parameters calculated on CT images, some of them have units and the others do not. We summarize their units as follows: 7 dimensionless measures (skewness and kurtosis of the GLHA parameters; entropy, angular second moment, correlation and maximum probability of the SGLDM parameters; busyness of the NGTDM parameters), 4 measures with the unit of HU (mean and standard deviation of the GLHA parameters; sum mean of the SGLDM parameters; strength of the NGTDM parameters), 4 measures with the unit of HU² (contrast, variance and cluster tendency of the SGLDM parameters; complexity of the NGTDM parameters), 1 measure with the unit of HU³ (contrast of the NGTDM parameters), 2 measures with the unit of 1/HU (homogeneity of the SGLDM parameters; coarseness of the NGTDM parameters) and 1 measure with the unit of 1/HU² (inverse difference moment of the SGLDM parameters).